Cloth folding is a complex task due to the inevitable self-occlusions of clothes, their complicated dynamics, and the disparate materials, geometries, and textures that garments can have. In this work, we learn folding actions conditioned on text commands. Translating high-level, abstract instructions into precise robotic actions requires sophisticated language understanding and manipulation capabilities. To do that, we leverage a pre-trained vision-language model and repurpose it to predict manipulation actions. Our model, BiFold, can take context into account and achieves state-of-the-art performance on an existing language-conditioned folding benchmark. To address the lack of annotated bimanual folding data, we introduce a novel dataset with automatically parsed actions and language-aligned instructions, enabling better learning of text-conditioned manipulation. BiFold attains the best performance on our dataset and demonstrates strong generalization to new instructions, garments, and environments.

Robotic-assisted dressing has the potential to significantly aid both patients as well as healthcare personnel, reducing the workload and improving the efficiency in clinical settings. While substantial progress has been made in robotic dressing assistance, prior works typically assume that garments are already unfolded and ready for use. However, in medical applications gowns and aprons are often stored in a folded configuration, requiring an additional unfolding step. In this paper, we introduce the pre-dressing step, the process of unfolding garments prior to assisted dressing. We leverage imitation learning for learning three manipulation primitives, including both high and low acceleration motions. In addition, we employ a visual classifier to categorise the garment state as closed, partly opened, and fully opened. We conduct an empirical evaluation of the learned manipulation primitives as well as their combinations. Our results show that highly dynamic motions are not effective for unfolding freshly unpacked garments, where the combination of motions can efficiently enhance the opening configuration.

Due to their flexibility and ease of use, Dynamical Movement Primitives (DMPs) are widely used in robotics applications and research. DMPs combine linear dynamical systems to achieve robustness to perturbations and adaptation to moving targets with non-linear function approximators to fit a wide range of demonstrated trajectories.We propose a novel DMP formulation with a generalized logistic function as a delayed goal system. This formulation inherently has low initial jerk, and generates the bell-shaped velocity profiles that are typical of human movement. As the novel formulation is more expressive, it is able to fit a wide range of human demonstrations well, also without a non-linear forcing term. We exploit this increased expressiveness by automating the fitting of the dynamical system parameters through opti-mization. Our experimental evaluation demonstrates that this optimization regularizes the forcing term, and improves the interpolation accuracy of parametric DMPs.

The advent of tactile sensors in robotics has sparked many ideas on how robots can leverage direct contact measurements of their environment interactions to improve manipulation tasks. An important line of research in this regard is grasp force control, which aims to manipulate objects safely by limiting the amount of force exerted on the object. While prior works have either hand-modeled their force controllers, employed model-based approaches, or not shown sim-to-real transfer, we propose a model-free deep reinforcement learning approach trained in simulation and then transferred to the robot without further fine-tuning. We, therefore, present a simulation environment that produces realistic normal forces, which we use to train continuous force control policies. A detailed evaluation shows that the learned policy performs similarly or better than a hand-crafted baseline. Ablation studies prove that the proposed inductive bias and domain randomization facilitate sim-to-real transfer. Code, models, and supplementary videos are available on https://sites.google.com/view/rl-force-ctrl

Estimating the pose of an object from a monocular image is a fundamental inverse problem in computer vision. Due to its ill-posed nature, solving this problem requires incorporating deformation priors. In practice, many materials do not perceptibly shrink or extend when manipulated, constituting a reliable and well-known prior. Mathematically, this translates to the preservation of the Riemannian metric. Neural networks offer the perfect playground to solve the surface reconstruction problem as they can approximate surfaces with arbitrary precision and allow the computation of differential geometry quantities. This paper presents an approach for inferring continuous deformable surfaces from a sequence of images, which is benchmarked against several techniques and achieves state-of-the-art performance without the need for offline training. Being a method that performs per-frame optimization, our method can refine its estimates, contrary to those based on performing a single inference step. Despite enforcing differential geometry constraints at each update, our approach is the fastest of all the tested optimization-based methods.

Teaching complex manipulation skills, such as folding garments, to a bi-manual robot is a very challenging task, which is often tackled through learning from demon-stration. The few datasets of garment-folding demonstrations available nowadays to the robotics research community have been either gathered from human demonstrations or generated through simulation. The former have the great difficulty of perceiving both cloth state and human action as well as transferring them to the dynamic control of the robot, while the latter require coding human motion into the simulator in open loop, i.e., without incorporating the visual feedback naturally used by people, resulting in far-from-realistic movements. In this article, we present an accurate dataset of human cloth folding demonstrations. The dataset is collected through our novel virtual reality (VR) framework, based on Unity's 3D platform and the use of an HTC Vive Pro system. The framework is capable of simulating realistic garments while allowing users to interact with them in real time through handheld controllers. By doing so, and thanks to the immersive experience, our framework permits exploiting human visual feedback in the demonstrations while at the same time getting rid of the difficulties of capturing the state of cloth, thus simplifying data acquisition and resulting in more realistic demonstrations. We create and make public a dataset of cloth manipulation sequences, whose cloth states are semantically labeled in an automatic way by using a novel low-dimensional cloth representation that yields a very good separation between different cloth configurations.

Robotic cloth manipulation is an increasingly relevant area of research, challenging classic control algorithms due to the deformable nature of cloth. While it is possible to apply linear model predictive control to make the robot move the cloth according to a given reference, this approach suffers from a large dimensionality of the state-space representation of the cloth models. To address this issue, in this work we study the application of an input-output model predictive control strategy, based on quadratic dynamic matrix control, to robotic cloth manipulation. To account for uncertain disturbances on the cloth's motion, we further extend the algorithm with suitable chance constraints. In extensive simulated experiments, involving disturbances and obstacle avoidance, we show that quadratic dynamic matrix control can be successfully applied in different cloth manipulation scenarios, with significant gains in optimization speed compared to standard model predictive control strategies. The experiments further demonstrate that the closed-loop model used by quadratic dynamic matrix control can be beneficial to the tracking accuracy, leading to improvements over the standard predictive control strategy. Moreover, a preliminary experiment on a real robot shows that quadratic dynamic matrix control can indeed be employed in real settings.

Enabling a robot to perform new tasks is a complex endeavor, usually beyond the reach of non-technical users. For this reason, research efforts that aim at empowering end-users to teach robots new abilities using intuitive modes of interaction are valuable. In this article, we present INtuitive PROgramming 2 (INPRO2), a learning framework that allows inferring planning actions from demonstrations given by a human teacher. INPRO2 operates in an assistive scenario, in which the robot may learn from a healthcare professional (a therapist or caregiver) new cognitive exercises that can be later administered to patients with cognitive impairment. INPRO2 features significant improvements over previous work, namely: (1) exploitation of negative examples; (2) proactive interaction with the teacher to ask questions about the legality of certain movements; and (3) learning goals in addition to legal actions. Through simulations, we show the performance of different proactive strategies for gathering negative examples. Real-world experiments with human teachers and a TIAGo robot are also presented to qualitatively illustrate INPRO2.

Abstract Socially assistive robots have the potential to augment and enhance therapist’s effectiveness in repetitive tasks such as cognitive therapies. However, their contribution has generally been limited as domain experts have not been fully involved in the entire pipeline of the design process as well as in the automatisation of the robots’ behaviour. In this article, we present aCtive leARning agEnt aSsiStive bEhaviouR (CARESSER), a novel framework that actively learns robotic assistive behaviour by leveraging the therapist’s expertise (knowledge-driven approach) and their demonstrations (data-driven approach). By exploiting that hybrid approach, the presented method enables in situ fast learning, in a fully autonomous fashion, of personalised patient-specific policies. With the purpose of evaluating our framework, we conducted two user studies in a daily care centre in which older adults affected by mild dementia and mild cognitive impairment ( N = 22) were requested to solve cognitive exercises with the support of a therapist and later on of a robot endowed with CARESSER. Results showed that: (i) the robot managed to keep the patients’ performance stable during the sessions even more so than the therapist; (ii) the assistance offered by the robot during the sessions eventually matched the therapist’s preferences. We conclude that CARESSER, with its stakeholder-centric design, can pave the way to new AI approaches that learn by leveraging human–human interactions along with human expertise, which has the benefits of speeding up the learning process, eliminating the need for the design of complex reward functions, and finally avoiding undesired states.

Abstract Task planning in human–robot environments tends to be particularly complex as it involves additional uncertainty introduced by the human user. Several plans, entailing few or various differences, can be obtained to solve the same given task. To choose among them, the usual least-cost plan criteria is not necessarily the best option, because here, human constraints and preferences come into play. Knowing these user preferences is very valuable to select an appropriate plan, but the preference values are usually hard to obtain. In this context, we propose the Space-of-Plans-based Suggestions (SoPS) algorithms that can provide suggestions for some planning predicates, which are used to define the state of the environment in a task planning problem where actions modify the predicates. We denote these predicates as suggestible predicates, of which user preferences are a particular case. The first algorithm is able to analyze the potential effect of the unknown predicates and provide suggestions to values for these unknown predicates that may produce better plans. The second algorithm is able to suggest changes to already known values that potentially improve the obtained reward. The proposed approach utilizes a Space of Plans Tree structure to represent a subset of the space of plans. The tree is traversed to find the predicates and the values that would most increase the reward, and output them as a suggestion to the user. Our evaluation in three preference-based assistive robotics domains shows how the proposed algorithms can improve task performance by suggesting the most effective predicate values first.

Over the last decade, the ability to teach actions to robots in a user-friendly way has gained relevance, and a practical way of teaching robots a new task is to use Inverse Reinforcement Learning (IRL). In IRL, an expert teacher shows the robot a desired behaviour and an agent builds a model of the reward. The agent can also infer a policy that performs in an optimal way within the limitations of the knowledge provided to it. However, most IRL approaches assume an (almost) optimal performance of the teaching agent, which might become unpractical if the teacher is not actually an expert. In addition, most IRL focus on discrete state-action spaces that limit their applicability to certain real-world problems such as within the context of direct Policy Search (PS) reinforcement learning. Therefore, in this paper we introduce Ordinal Inverse Reinforcement Learning (OrdIRL) for continuous state variables, in which the teacher can qualitatively evaluate robot performance by selecting one among the predefined performance levels (e.g. {bad, medium, good} for three tiers of performance). Once the OrdIRL has fit an ordinal distribution to the data, we propose to use Bayesian Optimization (BO) to either gain knowledge on the inferred model (exploration) or find a policy or action that maximizes the expected reward given the prior knowledge on the reward (exploitation). In the case of large-dimensional state-action spaces, we use Dimensionality Reduction (DR) techniques and perform the BO in the latent space. Experimental results on simulation and with a robot arm show how this approach allows for learning the reward function with small data.

The success of human-robot interaction is strongly affected by the people’s ability to infer others’ intentions and behaviours, and the level of people’s trust that others will abide by their same principles and social conventions to achieve a common goal. The ability of understanding and reasoning about other agents’ mental states is known as Theory of Mind (ToM). ToM and trust, therefore, are key factors in the positive outcome of human-robot interaction. We believe that a robot endowed with a ToM is able to gain people’s trust, even when this may occasionally make errors.In this work, we present a user study in the field in which participants (N=123) interacted with a robot that may or may not have a ToM, and may or may not exhibit erroneous behaviour. Our findings indicate that a robot with ToM is perceived as more reliable, and they trusted it more than a robot without a ToM even when the robot made errors. Finally, ToM results to be a key driver for tuning people’s trust in the robot even when the initial condition of the interaction changed (i.e., loss and regain of trust in a longer relationship).

Recognition in planning seeks to find agent intentions, goals or activities given a set of observations and a knowledge library (e.g. goal states, plans or domain theories). In this work we introduce the problem of Online Action Recognition. It consists in recognizing, in an open world, the planning action that best explains a partially observable state transition from a knowledge library of first-order STRIPS actions, which is initially empty. We frame this as an optimization problem, and propose two algorithms to address it: Action Unification (AU) and Online Action Recognition through Unification (OARU). The former builds on logic unification and generalizes two input actions using weighted partial MaxSAT. The latter looks for an action within the library that explains an observed transition. If there is such action, it generalizes it making use of AU, building in this way an AU hierarchy. Otherwise, OARU inserts a Trivial Grounded Action (TGA) in the library that explains just that transition. We report results on benchmarks from the International Planning Competition and PDDLGym, where OARU recognizes actions accurately with respect to expert knowledge, and shows real-time performance.

In this paper, we present a learning approach to facilitate the teaching of new board exercises to assistive robotic systems. We formulate the problem as the learning of action models using Boolean predicates, disjunctive preconditions, and existential quantifiers from demonstrations of successful exercise executions. To be able to cope with exercises whose rules depend on a set of features that are initialized at the beginning of each play-out, we introduce the concept of dynamic context. Furthermore, we show how the learnt knowledge can be represented intuitively in a graphical interface that helps the caregiver understand what the system has learnt. As validation, we conducted a user study in which we evaluated whether and to which extent different types of feedback can affect the subjects’ performance while teaching three types of exercises: (1) sorting numbers; (2) arranging letters; and (3) reproducing shapes sequences in reversed order. The results suggest that textual and graphical feedback are beneficial.

Assistive Robots have an inherent need of adapting to the user they are assisting. This is crucial for the correct development of the task, user safety, and comfort. However, adaptation can be performed in several manners. We believe user preferences are key to this adaptation. In this article, we evaluate the use of preferences for Physically Assistive Robotics tasks in a Human-Robot Interaction user evaluation. Three assistive tasks have been implemented consisting of assisted feeding, shoe-fitting, and jacket dressing, where the robot performs each task in a different manner based on user preferences. We assess the ability of the users to determine which execution of the task used their chosen preferences (if any). The obtained results show that most of the users were able to successfully guess the cases where their preferences were used even when they had not seen the task before. We also observe that their satisfaction with the task increases when the chosen preferences are employed. Finally, we also analyze the user’s opinions regarding assistive tasks and preferences, showing promising expectations as to the benefits of adapting the robot behavior to the user through preferences.

Robotic manipulators are reaching a state where we could see them in household environments in the following decade. Nevertheless, such robots need to be easy to instruct by lay people. This is why kinesthetic teaching has become very popular in recent years, in which the robot is taught a motion that is encoded as a parametric function - usually a Movement Primitive (MP)-. This approach produces trajectories that are usually suboptimal, and the robot needs to be able to improve them through trial-and-error. Such optimization is often done with Policy Search (PS) reinforcement learning, using a given reward function. PS algorithms can be classified as model-free, where neither the environment nor the reward function are modelled, or model-based, which can use a surrogate model of the reward function and/or a model for the dynamics of the task. However, MPs can become very high-dimensional in terms of parameters, which constitute the search space, so their optimization often requires too many samples. In this paper, we assume we have a robot motion task characterized with an MP of which we cannot model the dynamics. We build a surrogate model for the reward function, that maps an MP parameter latent space (obtained through a Mutual-information-weighted Gaussian Process Latent Variable Model) into a reward. While we do not model the task dynamics, using mutual information to shrink the task space makes it more consistent with the reward and so the policy improvement is faster in terms of sample efficiency.

In this paper we present a Deep Reinforcement Learning approach to solve dynamic cloth manipulation tasks. Differing from the case of rigid objects, we stress that the followed trajectory (including speed and acceleration) has a decisive influence on the final state of cloth, which can greatly vary even if the positions reached by the grasped points are the same. We explore how goal positions for non-grasped points can be attained through learning adequate trajectories for the grasped points. Our approach uses few demonstrations to improve control policy learning, and a sparse reward approach to avoid engineering complex reward functions. Since perception of textiles is challenging, we also study different state representations to assess the minimum observation space required for learning to succeed. Finally, we compare different combinations of control policy encodings, demonstrations, and sparse reward learning techniques, and show that our proposed approach can learn dynamic cloth manipulation in an efficient way, i.e., using a reduced observation space, a few demonstrations, and a sparse reward.

Human-robot interaction is one of the keys of assistive robots. Robots are expected to be compliant with people but at the same time correctly perform the tasks. In such applications, Cartesian impedance control is preferred over joint control, as the desired interaction and environmental feedback can be described more naturally, and the force to be exerted by the robot can be readily adjusted.This paper addresses the problem of controlling a robot arm in the operational space with variable stiffness so as to continuously adapt the force exerted in each phase of motion according to the precision requirements. Moreover, performing dimensionality reduction we can separate the degrees of freedom (DoF) relevant for the task from the redundant ones. The stiffness of the former can be adjusted constantly to achieve the required accuracy, while task-redundant DoF can be used to achieve other goals such as avoiding obstacles by moving in the directions where accuracy is not critical. The designed method is tested teaching the robot to give water to drink to a model of human head. Our empirical results demonstrate that the robot can learn precision requirements from demonstration. Furthermore, dimensionality reduction is proved to be useful to avoid obstacles.

In the next few years, the amount and variety of context-aware robotic manipulator applications is expected to increase significantly, especially in household environments. In such spaces, thanks to programming by demonstration, non-expert people will be able to teach robots how to perform specific tasks, for which the adaptation to the environment is imperative, for the sake of effectiveness and users safety. These robot motion learning procedures allow the encoding of such tasks by means of parameterized trajectory generators, usually a Movement Primitive (MP) conditioned on contextual variables. However, naively sampled solutions from these MPs are generally suboptimal/inefficient, according to a given reward function. Hence, Policy Search (PS) algorithms leverage the information of the experienced rewards to improve the robot performance over executions, even for new context configurations. Given the complexity of the aforementioned tasks, PS methods face the challenge of exploring in high-dimensional parameter search spaces. In this work, a solution combining Bayesian Optimization, a data-efficient PS algorithm, with covariate Gaussian Process Latent Variable Models, a recent Dimensionality Reduction technique, is presented. It enables reducing dimensionality and exploiting prior demonstrations to converge in few iterations, while also being compliant with context requirements. Thus, contextual variables are considered in the latent search space, from which a surrogate model for the reward function is built. Then, samples are generated in a low-dimensional latent space, and mapped to a context-dependent trajectory. This allows us to drastically reduce the search space with the covariate GPLVM, e.g. from 105 to 2 parameters, plus a few contextual features. Experimentation in two different scenarios proves the data-efficiency and the power of dimensionality reduction of our approach.

Learning is usually performed by observing real robot executions. Physics-based simulators are a good alternative for providing highly valuable information while avoiding costly and potentially destructive robot executions. We present a novel approach for learning the probabilities of symbolic robot action outcomes. This is done leveraging different environments, such as physics-based simulators, in execution time. To this end, we propose MENID (Multiple Environment Noise Indeterministic Deictic) rules, a novel representation able to cope with the inherent uncertainties present in robotic tasks. MENID rules explicitly represent each possible outcomes of an action, keep memory of the source of the experience, and maintain the probability of success of each outcome. We also introduce an algorithm to distribute actions among environments, based on previous experiences and expected gain. Before using physics-based simulations, we propose a methodology for evaluating different simulation settings and determining the least time-consuming model that could be used while still producing coherent results. We demonstrate the validity of the approach in a dismantling use case, using a simulation with reduced quality as simulated system, and a simulation with full resolution where we add noise to the trajectories and some physical parameters as a representation of the real system.

Robots are expected to become part of everyday life. However, while there have been important breakthroughs during the recent decades in terms of technological advances, the ability of robots to interact with humans intuitively and effectively is still an open challenge. In this paper, we aim to evaluate how humans interpret and leverage backchannel cues exhibited by a robot which interacts with them in an entertainment context. To do so, a conceptual model was designed to investigate the legibility and the effectiveness of a designed social cue, called SOCial ImmediAcy BackchanneL cuE (SOCIABLE), on participant's performance. In addition, user's attitude and cognitive capability were integrated into the model as an estimator of participants' motivation and ability to process the cue. In working toward such a goal, we conducted a two-day long user study (N=114) at an international event with untrained participants who were not aware of the social cue the robot was able to provide. The results showed that participants were able to perceive the social signal generated from SOCIABLE and thus, they benefited from it. Our findings provide some important insights for the design of effective and instantaneous backchannel cues and the methodology for evaluating them in social robots.

Compliant and soft hands have gained a lot of attention in the past decade because of their ability to adapt to the shape of the objects, increasing their effectiveness for grasping. However, when it comes to grasping highly flexible objects such as textiles, we face the dual problem: it is the object that will adapt to the shape of the hand or gripper. In this context, the classic grasp analysis or grasping taxonomies are not suitable for describing textile objects grasps. This article proposes a novel definition of textile object grasps that abstracts from the robotic embodiment or hand shape and recovers concepts from the early neuroscience literature on hand prehension skills. This framework enables us to identify what grasps have been used in literature until now to perform robotic cloth manipulation, and allows for a precise definition of all the tasks that have been tackled in terms of manipulation primitives based on regrasps. In addition, we also review what grippers have been used. Our analysis shows how the vast majority of cloth manipulations have relied only on one type of grasp, and at the same time we identify several tasks that need more variety of grasp types to be executed successfully. Our framework is generic, provides a classification of cloth manipulation primitives and can inspire gripper design and benchmark construction for cloth manipulation.

Collecting data in Human-Robot Interaction for training learning agents might be a hard task to accomplish. This is especially true when the target users are older adults with dementia since this usually requires hours of interactions and puts quite a lot of workload on the user. This paper addresses the problem of importing the Personas technique from HRI to create fictional patients' profiles. We propose a Persona-Behaviour Simulator tool that provides, with high-level abstraction, user's actions during an HRI task, and we apply it to cognitive training exercises for older adults with dementia. It consists of a Persona Definition that characterizes a patient along four dimensions and a Task Engine that provides information regarding the task complexity. We build a simulated environment where the high-level user's actions are provided by the simulator and the robot initial policy is learned using a Q-learning algorithm. The results show that the current simulator provides a reasonable initial policy for a defined Persona profile. Moreover, the learned robot assistance has proved to be robust to potential changes in the user's behaviour. In this way, we can speed up the fine-tuning of the rough policy during the real interactions to tailor the assistance to the given user. We believe the presented approach can be easily extended to account for other types of HRI tasks; for example, when input data is required to train a learning algorithm, but data collection is very expensive or unfeasible. We advocate that simulation is a convenient tool in these cases.

For a safe and successful daily living assistance, far from the highly controlled environment of a factory, robots should be able to adapt to ever-changing situations. Programming such a robot is a tedious process that requires expert knowledge. An alternative is to rely on a high-level planner, but the generic symbolic representations used are not well suited to particular robot executions. Contrarily, motion primitives encode robot motions in a way that can be easily adapted to different situations. This paper presents a combined framework that exploits the advantages of both approaches. The number of required symbolic states is reduced, as motion primitives provide “smart actions” that take the current state and cope online with variations. Symbolic actions can include interactions (e.g., ask and inform) that are difficult to demonstrate. We show that the proposed framework can adapt to the user preferences (in terms of robot speed and robot verbosity), can readjust the trajectories based on the user movements, and can handle unforeseen situations. Experiments are performed in a shoe-dressing scenario. This scenario is particularly interesting because it involves a sufficient number of actions, and the human-robot interaction requires the handling of user preferences and unexpected reactions.

In recent years the topic of combining motion and symbolic planning to perform complex tasks in the field of robotics has received a lot of attention. The underlying idea is to have access at once to the reasoning capabilities of a task planner and to the ability of the motion planner to verify that the plan is feasible from a physical and geometrical point of view. The present work describes a framework to perform manipulation tasks that require the use of two robotic manipulators. To do so we employ a Hierarchical Task Network (HTN) planner interleaved with geometric constraint verification. In this framework we also consider observation actions and handle noisy perceptions from a probabilistic perspective. These ideas are put into practice by means of an experimental set-up in which two Barrett WAM robots have to cooperatively solve a geometric puzzle. Our findings provide further evidence that considering explicitly physical constraints during task planning, rather than deferring their validation to the moment of execution, is advantageous in terms of execution time and breadth of situations that can be handled.

Interaction of socially assistive robots with users is based on social cues coming from different interaction modalities, such as speech or gestures. However, using all modalities at all times may be inefficient as it can overload the user with redundant information and increase the task completion time. Additionally, users may favor certain modalities over the other as a result of their disability or personal preference. In this paper, we propose an Adaptive Modality Selection (AMS) algorithm that chooses modalities depending on the state of the user and the environment, as well as user preferences. The variables that describe the environment and the user state are defined as resources, and we posit that modalities are successful if certain resources possess specific values during their use. Besides the resources, the proposed algorithm takes into account user preferences which it learns while interacting with users. We tested our algorithm in simulations, and we implemented it on a robotic system that provides cognitive training, specifically Sequential memory exercises. Experimental results show that it is possible to use only a subset of available modalities without compromising the interaction. Moreover, we see a trend for users to perform better when interacting with a system with implemented AMS algorithm.

Dynamic movement primitives (DMPs) are widely used as movement parametrization for learning robot trajectories, because of their linearity in the parameters, rescaling robustness, and continuity. However, when learning a movement with DMPs, a very large number of Gaussian approximations needs to be performed. Adding them up for all joints yields too many parameters to be explored when using reinforcement learning (RL), thus requiring a prohibitive number of experiments/simulations to converge to a solution with a (locally or globally) optimal reward. In this paper, we address the process of simultaneously learning a DMP-characterized robot motion and its underlying joint couplings through linear dimensionality reduction (DR), which will provide valuable qualitative information leading to a reduced and intuitive algebraic description of such motion. The results in the experimental section not only show that we can effectively perform DR on DMPs while learning, but we can also obtain better learning curves, as well as additional information about each motion: linear mappings relating joint values and some latent variables.

We present a three-level cognitive system in a learning by demonstration context. The system allows for learning and transfer on the sensorimotor level as well as the planning level. The fundamentally different data structures associated with these two levels are connected by an efficient mid-level representation based on so-called “semantic event chains.” We describe details of the representations and quantify the effect of the associated learning procedures for each level under different amounts of noise. Moreover, we demonstrate the performance of the overall system by three demonstrations that have been performed at a project review. The described system has a technical readiness level (TRL) of 4, which in an ongoing follow-up project will be raised to TRL 6.

Movement Primitives (MPs) have been widely used over the last years for learning robot motion tasks with direct Policy Search (PS) reinforcement learning. Among them, Probabilistic Movement Primitives (ProMPs) are a kind of MP based on a stochastic representation over sets of trajectories, which benefits from the properties of probability operations. However, the generation of such ProMPs requires a set of demonstrations to capture motion variability. Additionally, using context variables to modify trajectories coded as MPs is a popular approach nowadays in order to adapt motion to environmental variables. This paper proposes a contextual representation of ProMPs that allows for an easy adaptation to changing situations through context variables, by reparametrizing motion with them. Moreover, we propose a way of initializing contextual trajectories without the need of real robot demonstrations, by setting an initial position, a final position, and a number of trajectory interest points, where the contextual variables are evaluated. The parametrizations obtained show to be accurate while relieving the user from the need of performing costly computations such as conditioning. Additionally, using this contextual representation, we propose a simple yet effective quadratic optimization-based obstacle avoidance method for ProMPs. Experiments in simulation and on a real robot show the promise of the approach.

Assistive devices and technologies are getting common and some commercial products are starting to be available. However, the deployment of robots able to physically interact with a person in an assistive manner is still a challenging problem. Apart from the design and control, the robot must be able to adapt to the user it is attending in order to become a useful tool for caregivers. This robot behavior adaptation comes through the definition of user preferences for the task such that the robot can act in the user's desired way. This article presents a taxonomy of user preferences for assistive scenarios, including physical interactions, that may be used to improve robot decision-making algorithms. The taxonomy categorizes the preferences based on their semantics and possible uses. We propose the categorization in two levels of application (global and specific) as well as two types (primary and modifier). Examples of real preference classifications are presented in three assistive tasks: feeding, shoe fitting and coat dressing.

Probabilistic planners have improved recently to the point that they can solve difficult tasks with complex and expressive models. In contrast, learners cannot tackle yet the expressive models that planners do, which forces complex models to be mostly handcrafted. We propose a new learning approach that can learn relational probabilistic models with both action effects and exogenous effects. The proposed learning approach combines a multi-valued variant of inductive logic programming for the generation of candidate models, with an optimization method to select the best set of planning operators to model a problem. We also show how to combine this learner with reinforcement learning algorithms to solve complete problems. Finally, experimental validation is provided that shows improvements over previous work in both simulation and a robotic task. The robotic task involves a dynamic scenario with several agents where a manipulator robot has to clear the tableware on a table. We show that the exogenous effects learned by our approach allowed the robot to clear the table in a more efficient way.

Policy search (PS) algorithms are widely used for their simplicity and effectiveness in finding solutions for robotic problems. However, most current PS algorithms derive policies by statistically fitting the data from the best experiments only. This means that experiments yielding a poor performance are usually discarded or given too little influence on the policy update. In this paper, we propose a generalization of the relative entropy policy search (REPS) algorithm that takes bad experiences into consideration when computing a policy. The proposed approach, named dual REPS (DREPS) following the philosophical interpretation of the duality between good and bad, finds clusters of experimental data yielding a poor behavior and adds them to the optimization problem as a repulsive constraint. Thus, considering that there is a duality between good and bad data samples, both are taken into account in the stochastic search for a policy. Additionally, a cluster with the best samples may be included as an attractor to enforce faster convergence to a single optimal solution in multimodal problems. We first tested our proposed approach in a simulated reinforcement learning setting and found that DREPS considerably speeds up the learning process, especially during the early optimization steps and in cases where other approaches get trapped in between several alternative maxima. Further experiments in which a real robot had to learn a task with a multimodal reward function confirm the advantages of our proposed approach with respect to REPS.

Reliable object discovery in realistic indoor scenes is a necessity for many computer vision and service robot applications. In these scenes, semantic segmentation methods have made huge advances in recent years. Such methods can provide useful prior information for object discovery by removing false positives and by delineating object boundaries. We propose a novel method that combines bottom-up object discovery and semantic priors for producing generic object candidates in RGB-D images. We use a deep learning method for semantic segmentation to classify colour and depth superpixels into meaningful categories. Separately for each category, we use saliency to estimate the location and scale of objects, and superpixels to find their precise boundaries. Finally, object candidates of all categories are combined and ranked. We evaluate our approach on the NYU Depth V2 dataset and show that we outperform other state-of-the-art object discovery methods in terms of recall.

Robots are becoming safe and smart enough to work alongside people not only on manufacturing production lines, but also in spaces such as houses, museums, or hospitals. This can be significantly exploited in situations in which a human needs the help of another person to perform a task, because a robot may take the role of the helper. In this sense, a human and the robotic assistant may cooperatively carry out a variety of tasks, therefore requiring the robot to communicate with the person, understand his/her needs, and behave accordingly. To achieve this, we propose a framework for a user to teach a robot collaborative skills from demonstrations. We mainly focus on tasks involving physical contact with the user, in which not only position, but also force sensing and compliance become highly relevant. Specifically, we present an approach that combines probabilistic learning, dynamical systems, and stiffness estimation to encode the robot behavior along the task. Our method allows a robot to learn not only trajectory following skills, but also impedance behaviors. To show the functionality and flexibility of our approach, two different testbeds are used: a transportation task and a collaborative table assembly.

Learning motion tasks in a real environment with deformable objects requires not only a Reinforcement Learning (RL) algorithm, but also a good motion characterization, a preferably compliant robot controller, and an agent giving feedback for the rewards/costs in the RL algorithm. In this paper, we unify all these parts in a simple but effective way to properly learn safety-critical robotic tasks such as wrapping a scarf around the neck (so far, of a mannequin).

Modern plant phenotyping requires active sensing technologies and particular exploration strategies. This article proposes a new method for actively exploring a 3D region of space with the aim of localizing special areas of interest for manipulation tasks over plants. In our method, exploration is guided by a multi-layer occupancy grid map. This map, together with a multiple-view estimator and a maximum-information-gain gathering approach, incrementally provides a better understanding of the scene until a task termination criterion is reached. This approach is designed to be applicable for any task entailing 3D object exploration where some previous knowledge of its general shape is available. Its suitability is demonstrated here for an eye-in-hand arm configuration in a leaf probing application.

Task learning in robotics requires repeatedly executing the same actions in different states to learn the model of the task. However, in real-world domains, there are usually sequences of actions that, if executed, may produce unrecoverable errors (e.g. breaking an object). Robots should avoid repeating such errors when learning, and thus explore the state space in a more intelligent way. This requires identifying dangerous action effects to avoid including such actions in the generated plans, while at the same time enforcing that the learned models are complete enough for the planner not to fall into dead-ends. We thus propose a new learning method that allows a robot to reason about dead-ends and their causes. Some such causes may be dangerous action effects (i.e., leading to unrecoverable errors if the action were executed in the given state) so that the method allows the robot to skip the exploration of risky actions and guarantees the safety of planned actions. If a plan might lead to a dead-end (e.g., one that includes a dangerous action effect), the robot tries to find an alternative safe plan and, if not found, it actively asks a teacher whether the risky action should be executed. This method permits learning safe policies as well as minimizing unrecoverable errors during the learning process. Experimental validation of the approach is provided in two different scenarios: a robotic task and a simulated problem from the international planning competition. Our approach greatly increases success ratios in problems where previous approaches had high probabilities of failing.

Plant growth is a dynamic process, and the precise course of events during early plant development is of major interest for plant research. In this work, we investigate the growth of rosette plants by processing time-lapse videos of growing plants, where we use Nicotiana tabacum (tobacco) as a model plant. In each frame of the video sequences, potential leaves are detected using a leaf-shape model. These detections are prone to errors due to the complex shape of plants and their changing appearance in the image, depending on leaf movement, leaf growth, and illumination conditions. To cope with this problem, we employ a novel graph-based tracking algorithm which can bridge gaps in the sequence by linking leaf detections across a range of neighboring frames. We use the overlap of fitted leaf models as a pairwise similarity measure, and forbid graph edges that would link leaf detections within a single frame. We tested the method on a set of tobacco-plant growth sequences, and could track the first leaves of the plant, including partially or temporarily occluded ones, along complete sequences, demonstrating the applicability of the method to automatic plant growth analysis. All seedlings displayed approximately the same growth behavior, and a characteristic growth signature was found.

We propose a new approach for the segmentation of 3-D point clouds into geometric surfaces using adaptive surface models. Starting from an initial configuration, the algorithm converges to a stable segmentation through a new iterative split-and-merge procedure, which includes an adaptive mechanism for the creation and removal of segments. This allows the segmentation to adjust to changing input data along the movie, leading to stable, temporally coherent, and traceable segments. We tested the method on a large variety of data acquired with different range imaging devices, including a structured-light sensor and a time-of-flight camera, and successfully segmented the videos into surface segments. We further demonstrated the feasibility of the approach using quantitative evaluations based on ground-truth data.

Detecting objects in cluttered scenes is a necessary step for many robotic tasks and facilitates the interaction of the robot with its environment. Because of the availability of efficient 3D sensing devices as the Kinect, methods for the recognition of objects in 3D point clouds have gained importance during the last years. In this paper, we propose a new supervised learning approach for the recognition of objects from 3D point clouds using Conditional Random Fields, a type of discriminative, undirected probabilistic graphical model. The various features and contextual relations of the objects are described by the potential functions in the graph. Our method allows for learning and inference from unorganized point clouds of arbitrary sizes and shows significant benefit in terms of computational speed during prediction when compared to a state-of-the-art approach based on constrained optimization.

In this paper we present an automated system that is able to track and grasp a moving object within the workspace of a manipulator using range images acquired with a Microsoft Kinect sensor. Realtime tracking is achieved by a geometric particle filter on the affine group. Based on the tracked output, the pose of a 7-DoF WAM robotic arm is continuously updated using dynamic motor primitives until a distance measure between the tracked object and the gripper mounted on the arm is below a threshold. Then, it closes its three fingers and grasps the object. The tracker works in realtime and is robust to noise and partial occlusions. Using only the depth data makes our tracker independent of texture which is one of the key design goals in our approach. An experimental evaluation is provided along with a comparison of the proposed tracker with state-of-the-art approaches, including the OpenNI-tracker. The developed system is integrated with ROS and made available as part of IRI's ROS stack.

We describe a system allowing a robot to learn goal-directed manipulation sequences such as steps of an assembly task. Learning is based on a free mix of exploration and instruction by an external teacher, and may be active in the sense that the system tests actions to maximize learning progress and asks the teacher if needed. The main component is a symbolic planning engine that operates on learned rules, defined by actions and their pre- and postconditions. Learned by model-based reinforcement learning, rules are immediately available for planning. Thus, there are no distinct learning and application phases. We show how dynamic plans, replanned after every action if necessary, can be used for automatic execution of manipulation sequences, for monitoring of observed manipulation sequences, or a mix of the two, all while extending and refining the rule base on the fly. Quantitative results indicate fast convergence using few training examples, and highly effective teacher intervention at early stages of learning.

Dynamic Movement Primitives (DMP) are nowadays widely used as movement parametrization for learning trajectories, because of their linearity in the parameters, rescaling robustness and continuity. However, when learning a movement with a robot using DMP, many parameters may need to be tuned, requiring a prohibitive number of experiments/simulations to converge to a solution with a locally or globally optimal reward.

Research in learning from demonstration has focused on transferring movements from humans to robots. However, a need is arising for robots that do not just replicate the task on their own, but that also interact with humans in a safe and natural way to accomplish tasks cooperatively. Robots with variable impedance capabilities opens the door to new challenging applications, where the learning algorithms must be extended by encapsulating force and vision information. In this paper we propose a framework to transfer impedance-based behaviors to a torque-controlled robot by kinesthetic teaching. The proposed model encodes the examples as a task-parameterized statistical dynamical system, where the robot impedance is shaped by estimating virtual stiffness matrices from the set of demonstrations. A collaborative assembly task is used as testbed. The results show that the model can be used to modify the robot impedance along task execution to facilitate the collaboration, by triggering stiff and compliant behaviors in an on-line manner to adapt to the user's actions.

We introduce a novel approach to automatically recover 3D human pose from a single image. Most previous work follows a pipelined approach: initially, a set of 2D features such as edges, joints or silhouettes are detected in the image, and then these observations are used to infer the 3D pose. Solving these two problems separately may lead to erroneous 3D poses when the feature detector has performed poorly. In this paper, we address this issue by jointly solving both the 2D detection and the 3D inference problems. For this purpose, we propose a Bayesian framework that integrates a generative model based on latent variables and discriminative 2D part detectors based on HOGs, and perform inference using evolutionary algorithms. Real experimentation demonstrates competitive results, and the ability of our methodology to provide accurate 2D and 3D pose estimations even when the 2D detectors are inaccurate.

This paper presents a method to estimate external forces exerted on a manipulator during motion, avoiding the use of a sensor. The method is based on task-oriented dynamics model learning and a robust disturbance state observer. The combination of both leads to an efficient torque observer that can be incorporated to any control scheme. The use of a learning-based approach avoids the need of analytical models of joints' friction or Coriolis dynamics effects.

Most current depth sensors provide 2.5D range images in which depth values are assigned to a rectangular 2D array. In this paper we take advantage of this structured information to build an efficient shape descriptor which is about two orders of magnitude faster than competing approaches, while showing similar performance in several tasks involving deformable object recognition. Given a 2D patch surrounding a point and its associated depth values, we build the descriptor for that point, based on the cumulative distances between their normals and a discrete set of normal directions. This processing is made very efficient using integral images, even allowing to compute descriptors for every range image pixel in a few seconds. The discriminative power of our descriptor, dubbed FINDDD, is evaluated in three different scenarios: recognition of specific cloth wrinkles, instance recognition from geometry alone, and detection of reliable and informed grasping points.

Vision-based robot localization outdoors has remained more elusive than its indoors counterpart. Drastic illumination changes and the scarceness of suitable landmarks are the main difficulties. This paper attempts to surmount them by deviating from the main trend of using local features. Instead, a global descriptor called landmark-view is defined, which aggregates the most visually-salient landmarks present in each scene. Thus, landmark co-occurrence and spatial and saliency relationships between them are added to the single landmark characterization, based on saliency and color distribution. A suitable framework to compare landmark-views is developed, and it is shown how this remarkably enhances the recognition performance, compared against single landmark recognition. A view-matching model is constructed using logistic regression. Experimentation using 45 views, acquired outdoors, containing 273 landmarks, yielded good recognition results. The overall percentage of correct view classification obtained was 80.6%, indicating the adequacy of the approach.

Markerless 3D human pose detection from a single image is a severely underconstrained problem because different 3D poses can have similar image projections. In order to handle this ambiguity, current approaches rely on prior shape models that can only be correctly adjusted if 2D image features are accurately detected. Unfortunately, although current 2D part detector algorithms have shown promising results, they are not yet accurate enough to guarantee a complete disambiguation of the 3D inferred shape. In this paper, we introduce a novel approach for estimating 3D human pose even when observations are noisy. We propose a stochastic sampling strategy to propagate the noise from the image plane to the shape space. This provides a set of ambiguous 3D shapes, which are virtually undistinguishable from their image projections. Disambiguation is then achieved by imposing kinematic constraints that guarantee the resulting pose resembles a 3D human shape. We validate the method on a variety of situations in which state-of-the-art 2D detectors yield either inaccurate estimations or partly miss some of the body parts.

Detecting grasping points is a key problem in cloth manipulation. Most current approaches follow a multiple re-grasp strategy for this purpose, in which clothes are sequentially grasped from different points until one of them yields to a desired configuration. In this paper, by contrast, we circumvent the need for multiple re-graspings by building a robust detector that identifies the grasping points, generally in one single step, even when clothes are highly wrinkled. In order to handle the large variability a deformed cloth may have, we build a Bag of Features based detector that combines appearance and 3D geometry features. An image is scanned using a sliding window with a linear classifier, and the candidate windows are refined using a non-linear SVM and a “grasp goodness” criterion to select the best grasping point. We demonstrate our approach detecting collars in deformed polo shirts, using a Kinect camera. Experimental results show a good performance of the proposed method not only in identifying the same trained textile object part under severe deformations and occlusions, but also the corresponding part in other clothes, exhibiting a degree of generalization.

Motivated by the need of a robust and practical Inverse Kinematics (IK) algorithm for the WAM robot arm, we reviewed the most used closed-loop methods for redundant robots, analysing their main points of concern: convergence, numerical error, singularity handling, joint limit avoidance, and the capability of reaching secondary goals. As a result of the experimental comparison, we propose two enhancements. The first is to filter the singular values of the Jacobian matrix before calculating its pseudoinverse in order to obtain a more numerically robust result. The second is to combine a continuous task priority strategy with selective damping to generate smoother trajectories. Experimentation on the WAM robot arm shows that these two enhancements yield an IK algorithm that improves on the reviewed state-of-the-art ones, in terms of the good compromise it achieves between time step length, Jacobian conditioning, multiple task performance, and computational time, thus constituting a very solid option in practice. This proposal is general and applicable to other redundant robots.

Rigid object manipulation with robots has mainly relied on precise, expensive models and deterministic sequences. Given the great complexity of accurately modeling deformable objects, their manipulation seems to call for a rather different approach. This paper proposes a probabilistic planner, based on a Partially Observable Markov Decision Process (POMDP), targeted at reducing the inherent uncertainty of deformable object sorting. It is shown that a small set of unreliable actions and inaccurate perceptions suffices to accomplish the task, provided faithful statistics on both of them are collected beforehand. The planner has been applied to a clothes sorting task in a real case context with a depth and color sensor and a robotic arm. Experimental results show the promise of the approach since more than 95% certainty of having isolated a piece of clothing is reached in an average of four steps for quite entangled initial clothing configurations.

The calibration of serial manipulators with high numbers of degrees of freedom by means of machine learning is a complex and time-consuming task. With the help of a simple strategy, this complexity can be drastically reduced and the speed of the learning procedure can be increased. When the robot is virtually divided into shorter kinematic chains, these subchains can be learned separately and hence much more efficiently than the complete kinematics. Such decompositions, however, require either the possibility to capture the poses of all end effectors of all subchains at the same time, or they are limited to robots that fulfill special constraints. In this paper, an alternative decomposition is presented that does not suffer from these limitations. An offline training algorithm is provided in which the composite subchains are learned sequentially with dedicated movements. A second training scheme is provided to train composite chains simultaneously and online. Both schemes can be used together with many machine learning algorithms. In the simulations, an algorithm using parameterized self-organizing maps modified for online learning and Gaussian mixture models (GMMs) were chosen to show the correctness of the approach. The experimental results show that, using a twofold decomposition, the number of samples required to reach a given precision is reduced to twice the square root of the original number.

The kinematics of a robot with many degrees of freedom is a very complex function. Learning this function for a large workspace with a good precision requires a huge number of training samples, i.e., robot movements. In this paper, we introduce the Kinematic Bézier Map (KB-Map), a parameterizable model without the generality of other systems but whose structure readily incorporates some of the geometric constraints of a kinematic function. In this way, the number of training samples required is drastically reduced. Moreover, the simplicity of the model reduces learning to solving a linear least squares problem. Systematic experiments have been carried out showing the excellent interpolation and extrapolation capabilities of KB-Maps and their relatively low sensitivity to noise.

Supervision of long-lasting extensive botanic experiments is a promising robotic application that some recent technological advances have made feasible. Plant modelling for this application has strong demands, particularly in what concerns 3D information gathering and speed. This paper shows that Time-of-Flight (ToF) cameras achieve a good compromise between both demands, providing a suitable complement to color vision. A new method is proposed to segment plant images into their composite surface patches by combining hierarchical color segmentation with quadratic surface fitting using ToF depth data. Experimentation shows that the interpolated depth maps derived from the obtained surfaces fit well the original scenes. Moreover, candidate leaves to be approached by a measuring instrument are ranked, and then robot-mounted cameras move closer to them to validate their suitability to being sampled. Some ambiguities arising from leaves overlap or occlusions are cleared up in this way. The work is a proof-of-concept that dense color data combined with sparse depth as provided by a ToF camera yields a good enough 3D approximation for automated plant measuring at the high throughput imposed by the application.

We present a new method for segmenting color images into their composite surfaces by combining color segmentation with model-based fitting utilizing sparse depth data, acquired using time-of-flight (Swissranger, PMD CamCube) and stereo techniques. The main target of our work is the segmentation of plant structures, i.e., leaves, from color-depth images, and the extraction of color and 3D shape information for automating manipulation tasks. Since segmentation is performed in the dense color space, even sparse, incomplete, or noisy depth information can be used. This kind of data often represents a major challenge for methods operating in the 3D data space directly. To achieve our goal, we construct a three-stage segmentation hierarchy by segmenting the color image with different resolutions-assuming that “true” surface boundaries must appear at some point along the segmentation hierarchy. 3D surfaces are then fitted to the color-segment areas using depth data. Those segments which minimize the fitting error are selected and used to construct a new segmentation. Then, an additional region merging and a growing stage are applied to avoid over-segmentation and label previously unclustered points. Experimental results demonstrate that the method is successful in segmenting a variety of domestic objects and plants into quadratic surfaces. At the end of the procedure, the sparse depth data is completed using the extracted surface models, resulting in dense depth maps. For stereo, the resulting disparity maps are compared with ground truth and the average error is computed.

A 5-SPU robot with collinear universal joints is well suited to handle an axisymmetric tool, since it has five controllable degrees of freedom, and the remaining one is a free rotation around the tool. The kinematics of such a robot also having coplanar spherical joints has previously been studied as a rigid subassembly of a Stewart-Gough platform, which has been denoted a line-plane component. Here, we investigate how to move the leg attachments in the base and the platform without altering the robot's singularity locus. By introducing the so-called 3-D space of leg attachments, we prove that there are only three general topologies for the singularity locus corresponding to the families of quartically, cubically, and quadratically solvable 5-SPU robots. The members of the last family have only four assembly modes, which are obtained by solving two quadratic equations. Two practical features of these quadratically solvable robots are the large manipulability within each connected component and the fact that, for a fixed orientation of the tool, the singularity locus reduces to a plane.

Time-of-Flight (ToF) cameras deliver 3D images at 25 fps, offering great potential for developing fast object modeling algorithms. Surprisingly, this potential has not been extensively exploited up to now. A reason for this is that, since the acquired depth images are noisy, most of the available registration algorithms are hardly applicable. A further difficulty is that the transformations between views are in general not accurately known, a circumstance that multi-view object modeling algorithms do not handle properly under noisy conditions. In this work, we take into account both uncertainty sources (in images and camera poses) to generate spatially consistent 3D object models fusing multiple views with a probabilistic approach. We propose a method to compute the covariance of the registration process, and apply an iterative state estimation method to build object models under noisy conditions.

A 5-SPU robot with collinear universal joints is well suited to handling an axisymmetric tool, since it has 5 controllable DoFs and the remaining one is a free rotation around the tool. The kinematics of such a robot having also coplanar spherical joints has previously been studied as a rigid subassembly of a Stewart-Gough platform, it being denoted a line-plane component. It was shown that this component has 8 assembly modes corresponding to the roots of a bi-quartic polynomial. Here we identify a whole family of these 5-SPU robots having only 4 assembly modes, which are obtained by solving two quadratic equations. This family is defined by a simple proportionality constraint relating the coordinates of the base and platform attachments. A geometric interpretation of the architectural singularities of this type of robots in terms of conics is provided, which facilitates their avoidance at the design stage. Parallel singularities obey also a neat geometric structure, which permits deriving a cell decomposition of configuration space. Two practical features of these quadratically-solvable robots are the large maneuverability within each connected component and the fact that, for a fixed orientation of the tool, the singularity locus reduces to a plane.

Symmetries in discrete constraint satisfaction problems have been explored and exploited in the last years, but symmetries in continuous constraint problems have not received the same attention. Here we focus on permutations of the variables consisting of one single cycle. We propose a procedure that takes advantage of these symmetries by interacting with a continuous constraint solver without interfering with it. A key concept in this procedure are the classes of symmetric boxes formed by bisecting a n-dimensional cube at the same point in all dimensions at the same time. We analyze these classes and quantify them as a function of the cube dimensionality. Moreover, we propose a simple algorithm to generate the representatives of all these classes for any number of variables at very high rates. A problem example from the chemical field and the cyclic n-roots problem are used to show the performance of the approach in practice.

There are only three 6-SPS parallel manipulators with triangular base and platform, i.e., the octahedral, the flagged, and the partially flagged, which are studied in this paper. The forward kinematics of the octahedral manipulator is algebraically intricate, while those of the other two can be solved by three trilaterations. As an additional nice feature, the flagged manipulator is the only parallel platform for which a cell decomposition of its singularity locus has been derived. Here, we prove that the partially flagged manipulator also admits a well-behaved decomposition, technically called a stratification, some of whose strata are not topological cells, however. Remarkably, the adjacency diagram of the 5-D and 6-D strata (which shows what 5-D strata are contained in the closure of a 6-D one) is the same as for the flagged manipulator. The availability of such a decomposition permits devising a redundant 7-SPS manipulator, combining two partially flagged ones, which admits a control strategy that completely avoids singularities. Simulation results support these claims.

Any set of two legs in a Gough–Stewart platform sharing an attachment is defined as a$\Delta$component. This component links a point in the platform (base) to a line in the base (platform). Thus, if the two legs, which are involved in a$\Delta$component, are rearranged without altering the location of the line and the point in their base and platform local reference frames, the singularity locus of the Gough–Stewart platform remains the same, provided that no architectural singularities are introduced. Such leg rearrangements are defined as$\Delta$-transforms, and they can be applied sequentially and simultaneously. Although it may seem counterintuitive at first glance, the rearrangement of legs using simultaneous$\Delta$-transforms does not necessarily lead to leg configurations containing a$\Delta$component. As a consequence, the application of$\Delta$-transforms reveals itself as a simple, yet powerful, technique for the kinematic analysis of large families of Gough–Stewart platforms. It is also shown that these transforms shed new light on the characterization of architectural singularities and their associated self-motions.

Flagged manipulators are of interest because they are the only Stewart-Gough platforms for which a cell decomposition of their singularity loci is available. Here we show that the known family of such manipulators can be enlarged if one allows robot designs that, for some particular parameter values, become architecturally singular. Along this line, the most general 6-6 flagged manipulator is derived by applying a singularity-preserving transformation that leaves the relative position between two lines invariant. This transformation opens up the possibility of an "equal cross ratios" architectural singularity, which is shown to appear clearly in the factorization of the Jacobian determinant. From the 6-6 flagged manipulator, all the extended family of (possibly architecturally-singular) flagged manipulators is derived.

We propose a technique to speedup the learning of the inverse kinematics of a robot manipulator by decomposing it into two or more virtual robot arms. Unlike previous decomposition approaches, this one does not place any requirement on the robot architecture, and thus, it is completely general. Parametrized self-organizing maps are particularly adequate for this type of learning, and permit comparing results directly obtained and through the decomposition. Experimentation shows that time reductions of up to two orders of magnitude are easily attained.

Robot egomotion can be estimated from an acquired video stream up to the scale of the scene. To remove this uncertainty (and obtain true egomotion), a distance within the scene needs to be known. If no a priori knowledge on the scene is assumed, the usual solution is to derive "in some way" the initial distance from the camera to a target object. This paper proposes a new, very simple way to obtain such a distance, when a zooming camera is available and there is a planar target in the scene. Similarly to "two-grid calibration" algorithms, no estimation of the camera parameters is required, and no assumption on the optical axis stability between the different focal lengths is needed. Quite the reverse, the non stability of the optical axis between the different focal lengths is the key ingredient that enables to derive our depth estimate, by applying a result in projective geometry. Experiments carried out on a mobile robot platform show the promise of the approach.

This paper describes the extraction of visual landmarks from outdoor images for mobile robot applications. The concept of group of landmarks, called landmark-view, is introduced, aggregating the most relevant landmarks present in each scene. The relevance of the landmarks is determined by their relative visual saliency. Thus, landmark co-occurrence and spatial and saliency relationships between them are added to the single landmark descriptors, which are based on saliency and color distribution in chromaticity space. A suitable framework to compare landmark-views is developed, and it is shown how this remarkably enhances the recognition performance, compared against the single landmark recognition. A view-matching model is constructed using logistic regression. Experimentation using 45 views, acquired outdoors, containing 273 landmarks, yielded good recognition results. Of the 42 corresponding view pairs, 30 were recognized correctly, resulting in 71.4% of correct classification of similar views. Of the 948 non-corresponding view pairs, 768 were recognized correctly, resulting in 81.0% of correct classification in non-similar views. The overall percentage of correct view classification obtained was 80.6%, indicating the convenience of the approach.

The conditions for a parallel manipulator to be flagged can be simply expressed in terms of linear dependencies between the coordinates of its leg attachments, both on the base and on the platform. These dependencies permit to describe the manipulator singularities in terms of incidences between two flags (hence, the name ldquoflaggedrdquo). Although these linear dependencies might look, at first glance, too restrictive, in this paper, the family of flagged manipulators is shown to contain large subfamilies of six-legged and three-legged manipulators. The main interest of flagged parallel manipulators is that their singularity loci admit a well-behaved decomposition with a unique topology irrespective of the metrics of each particular design. In this paper, this topology is formally derived and all the cells, in the configuration space of the platform, of dimension 6 (nonsingular) and dimension 5 (singular), together with their adjacencies, are worked out in detail.

Flagged in-parallel manipulators are attractive because their singularity loci admit a well-behaved decomposition, with a unique topology irrespective of the metrics of each particular design. In this paper, this topology is formally derived and all the cells, in the configuration space of the platform, of dimension 6 (non-singular) and dimension 5 (singular), together with their adjacencies, are worked out in detail. This characterization of the singularity loci is useful to come up with designs which admit control strategies free of singularities. In particular, it is shown that by adding an extra leg to any flagged manipulator, the resulting 7-leg structure admits a control strategy (by appropriately choosing which leg remains passive) that completely avoids singularities

Some in-parallel robots, such as the 3-2-1 and the 3/2 manipulators, have attracted attention because their forward kinematics can be solved by three consecutive trilaterations. In this paper, we identify a class of these robots, which we call flagged manipulators, whose singularity loci admit a well-behaved decomposition, i.e., a stratification, derived from that of the flag manifold. Two remarkable properties must be highlighted. First, the decomposition has the same topology for all members in the class, irrespective of the metric details of each particular robot instance. Thus, we provide explicitly all the singular strata and their connectivity, which apply to all flagged manipulators without any tailoring. Second, the strata can be easily characterized geometrically, because it is possible to assign local coordinates to each stratum (in the configuration space of the manipulator) that correspond to uncoupled rotations and/or translations in the workspace.

The fully flexible navigation of autonomous vehicles in industrial environments is still unsolved. It is hard to conciliate strict precision requirements with quick adaptivity to new settings without undergoing costly rearrangements. We are pursuing a research project trying to combine the precision of laser-based local positioning with the flexibility of vision-based robot motion estimation. An enhanced circle approach to dynamic triangulation combining laser and odometric signals has been used to improve positioning accuracy. As regards to vision, a novel technique relating the deformation of contours in an image sequence to the 3D motion underwent by the camera has been developed. Interestingly, contours are fitted to objects already present in the environment, without requiring any presetting. In this paper, we describe a practical experience conducted in the warehouse of a beer production factory in Barcelona. A database containing the laser readings, image sequences and robot odometry along several trajectories was compiled, and subsequently processed off-line in order to assess the accuracies of both techniques under a variety of circumstances. In all, vision-based estimation turned out to be about one order of magnitude less precise than laser-based positioning, which qualifies the vision-based technique as a promising alternative to accomplish robot transfers across long distances, such as those needed in a warehouse, while backing up on laser-based positioning when accurate docking for loading and unloading operations is needed.

The main drawback of using neural networks or other example-based learning procedures to approximate the inverse kinematics (IK) of robot arms is the high number of training samples (i.e., robot movements) required to attain an acceptable precision. We propose here a trick, valid for most industrial robots, that greatly reduces the number of movements needed to learn or relearn the IK to a given accuracy. This trick consists in expressing the IK as a composition of learnable functions, each having half the dimensionality of the original mapping. Off-line and on-line training schemes to learn these component functions are also proposed. Experimental results obtained by using nearest neighbors and parameterized self-organizing map, with and without the decomposition, show that the time savings granted by the proposed scheme grow polynomially with the precision required.

Given some geometric elements such as points and lines in R/sup 3/, subject to a set of pairwise distance constraints, the problem tackled in this paper is that of finding all possible configurations of these elements that satisfy the constraints. Many problems in robotics (such as the position analysis of serial and parallel manipulators) and CAD/CAM (such as the interactive placement of objects) can be formulated in this way. The strategy herein proposed consists of looking for some of the a priori unknown distances, whose derivation permits solving the problem rather trivially. Finding these distances relies on a branch-and-prune technique, which iteratively eliminates from the space of distances entire regions which cannot contain any solution. This elimination is accomplished by applying redundant necessary conditions derived from the theory of distance geometry. The experimental results qualify this approach as a promising one.

Given a set of affine varieties in R/sup 3/, i.e. planes, lines, and points, the problem tackled in this paper is that of finding all possible configurations for these varieties that satisfy a set of pairwise euclidean distances between them. Many problems in robotics - such as the forward kinematics of patroller manipulators or the contact formation problem between polyhedral models - can be formulated in this way. We propose herein a strategy that consists in finding some distances, that are unknown a priori, and whose derivation permits solving the problem rather trivially. Finding these distances relies on a branch-and-prune technique that iteratively eliminates from the space of distances entire regions which cannot contain any solution. The elimination is accomplished by applying redundant necessary conditions derived from the theory of Cayley-Menger determinants. The experimental results obtained qualify this approach as a promising one.

A constraint satisfaction problem, namely the generation of Balanced Incomplete Block Designs (v, b, r, kappa, lambda)-BIBDs, is cast in terms of function optimization. A family of cost functions that both suit the problem and admit a neural implementation is defined. An experimental comparison spanning this repertoire of cost functions and three neural relaxation strategies (Down-Hill search, Simulated Annealing and a new Parallel Mean Search procedure), as applied to all BIBDs of up to 1000 entries, has been undertaken. The experiments were performed on a Connection Machine CM-200 and their analysis required a careful study of performance measures. The simplest cost function stood out as the best one for the three strategies. Parallel Mean Search, with several processors searching cooperatively in parallel, could solve a larger number of problems than the same number of processors working independently, but Simulated Annealing yielded overall the best results. Other conclusions, as detailed in the paper, could be drawn from the comparison, BIBDs remaining a challenging problem for neural optimization algorithms.

We show that minimizing the expected error of a feedforward network over a distribution of weights results in an approximation that tends to be independent of network size as the number of hidden units grows. This minimization can be easily performed, and the complexity of the resulting function implemented by the network is regulated by the variance of the weight distribution. For a fixed variance, there is a number of hidden units above which either the implemented function does not change or the change is slight and tends to zero as the size of the network grows. In sum, the control of the complexity depends on only the variance, not the architecture, provided it is large enough.

An active contour is used to track a target in a sequence recorded by a walking robot in an unstructured scene. The deformations of the contour are analysed in order to extract the robot's egomotion, from which we compute the epipolar geometry that guides the matching between different views of the scene. The results prove that the proposed solution is a promising alternative to the prevalent techniques based on the costly computation of displacement or velocity fields.

This work presents a landmark detection system for the walking robot operating in unknown unstructured outdoor environments. Most landmark detection approaches are not adequate for this application, since they rely on either structured information or a priori knowledge about the landmarks. Instead, the proposed system makes use of visual saliency concepts stemming from studies of animal and human perception. Thus, biologically inspired opponent features (in color and orientation) are searched for at different resolution levels. The implementation does not try to mimic nature, but rather to be as computationally efficient as possible. Thus, salient image regions ranging from relatively small to big sizes are detected using multiscale comparison techniques, based on pyramidal filtering. The experimental results obtained show that visual saliency permits detecting reliable natural landmarks without a priori knowledge about their characteristics or location.

The limited adaptivity of current robots is preventing their widespread application. However, nowadays there are mature techniques available to palliate this deficiency. After briefly surveying the several levels of adaptivity required and the disciplines addressing each of them, the paper concentrates on the contributions of the field of neural networks to improve sensorimotor adaptivity. Since sensorimotor mappings lie at the base of all robot activity, making them adaptable to the robot conditions (e.g., tear-and-wear) and environmental variations greatly widens the range of applications. Several experimental systems are described which rely on the following adaptive mappings: inverse kinematics, inverse dynamics, visuomotor and force-control mappings. Finally, some methodologic and computational issues are discussed.

Nonconvex polyhedral models of workpieces or robot parts can be directly tested for interference, without resorting to a previous decomposition into convex entities. We show that this interference detection, based on the elemental edge face intersection test, can be performed efficiently: a strategy based on applicability constraints reduces drastically the set of edge - face pairings that have to be considered for intersection. This is accomplished by using an appropriate representation, the spherical face orientation graph, developed by the authors, as well as feature pairing algorithms based on the plane sweep paradigm that have been adapted to work on that representation. Furthermore, the benefits of such a strategy extend to the computation of a lower distance bound between the polyhedra, both lowering the computational effort and improving the quality of the bound. Experimental results confirm the expected advantages of this strategy.

The authors and Venaille (1996) developed a prototype for visual robot positioning, based on global image descriptors and neural networks. Now, a procedure to automatically select subsets of image features most relevant to determine pose variations along each of the six degrees of freedom (DOFs) has been incorporated into the prototype. This procedure is based on a statistical measure of variable interdependence, called mutual information. Three families of features are considered in this paper: geometric moments, eigenfeatures and pose-image covariance vectors. The experimental results described show the quantitative and qualitative benefits of carrying out this feature selection prior to training the neural network: fewer network inputs need to be considered, thus considerably shortening training times; the DOFs that would yield larger errors can be determined beforehand, so that more informative features can be looked for; the ordering of the features selected for each DOF often admits a very natural interpretation, which in turn helps to provide insights for devising features tailored to each DOF.

We present a neural-network method to recalibrate automatically a commercial robot after undergoing wear or damage, which works on top of the nominal inverse kinematics embedded in its controller. Our starting point has been the work of Ritter et al. (1989, 1992) on the use of extended self-organizing maps to learn the whole inverse kinematics mapping from scratch. Besides adapting their approach to learning only the deviations from the nominal kinematics, we have introduced several modifications to improve the cooperation between neurons. These modifications not only speed up learning by two orders of magnitude, but also produce some desirable side effects, like parameter stability. After extensive experimentation through simulation, the recalibration system has been installed in the REIS robot included in the space-station mock-up at Daimler-Benz Aerospace. Tests performed in this set-up have been constrained by the need to preserve robot integrity, but the results have been concordant with those predicted through simulation.

A classical paradigm for interference detection between polyhedra consists in testing all edges of one polyhedron against all faces of the other one for intersection. If the relative orientation of the polyhedra is fixed, only certain edge-face pairs can intersect first, when the polyhedra come into contact. These candidate pairs are efficiently determined using a representation which the authors call spherical face orientation graph. By applying the interference test to candidates only, the computational effort is significantly reduced, as shown by experimental results with convex polyhedra. In the non-convex case, the strategy is conservative, but it still leads to savings.

Dealing with nonstationary processes requires quick adaptation while at the same time avoiding catastrophic forgetting. A neural learning technique that satisfies these requirements, without sacrificing the benefits of distributed representations, is presented. It relies on a formalization of the problem as the minimization of the error over the previously learned input-output patterns, subject to the constraint of perfect encoding of the new pattern. Then this constrained optimization problem is transformed into an unconstrained one with hidden-unit activations as variables. This new formulation leads to an algorithm for solving the problem, which we call learning with minimal degradation (LMD). Some experimental comparisons of the performance of LMD with backpropagation are provided which, besides showing the advantages of using LMD, reveal the dependence of forgetting on the learning rate in backpropagation. We also explain why overtraining affects forgetting and fault tolerance, which are seen as related problems.

Exact interference checking between two arbitrary polyhedra is known to have O(mn) complexity, where m and n are the number of edges in the two polyhedra. This is just a worst-case bound that still leaves plenty of room for algorithm improvement in practice. The algorithm presented herein has been developed so as to: 1. Minimize the number of operations that each pairing of edges entails. We prove that this factor is 4.5 for multiplications and 8.5 for additions. 2. Avoid the construction of auxiliary geometric entities. The standard approach is to decompose the nonconvex polyhedra (or their faces) into convex entities and then check for interference in this convex setting. This entails the construction of many fictitious edges and faces, which indirectly contribute to the growth of the complexity. 3. Permit the straightforward application of prunning strategies to most practical situations, so that the worst-case bound above is reached only when truly needed. 4 Allow the derivation of both directional and undirectional distance bounds between the polyhedra, which prove extremely useful for collision avoidance and local path planning. The simplicity and homogeneity of the algorithm has led to a quick implementation, which has been proven to be robust and fast. Some performance measurements are reported.>

Proposes a reinforcement learning architecture that allows an autonomous robot to acquire efficient navigation strategies in a few trials. Besides fast learning, the architecture has 3 further appealing features. (1) Since it learns from built-in reflexes, the robot is operational from the very beginning. (2) The robot improves its performance incrementally as it interacts with an initially unknown environment, and it ends up learning to avoid collisions even if its sensors cannot detect the obstacles. This is a definite advantage over non-learning reactive robots. (3) The robot exhibits high tolerance to noisy sensory data and good generalization abilities. All these features make this learning robot's architecture very well suited to real-world applications. The authors report experimental results obtained with a real mobile robot in an indoor environment that demonstrate the feasibility of this approach.>

The authors treat two different problems in the analysis of assemblies, and algorithms are described for each. The first problem is the selection of consistent sets of part feature relationships, and it corresponds to a search over possible configurations of parts that are consistent with feature set mappings. The second problem is the evaluation of the kinematic consistency of an assembly that has been defined by consistent feature sets. These two problems are linked together as two of the steps required in a search for all correct assembly configurations of a given set of parts. Several of the other necessary steps related to part interference, path feasibility, and workcell device kinematics are referred to but not analyzed. The proposed search algorithm is based on a constraint posting strategy, i.e. rather than generating and testing all specific alternatives, chunks of the search space are progressively removed from consideration by constraints that rule them out until one satisfactory alternative is found.>

The authors propose a robotic assembly system that is based on the least-commitment principle, in which the successive stages proceeding from planning to execution become progressively more specific. The system is partitioned into an offline automatic programming system and an online execution manager. The framework and objectives are presented, followed by an overview of the global system. The submodules of a system that has actually executed a simple pick/place task are described. Some conclusions and perspectives are outlined.>

A system for the automatic synthesis of assembly configuration is presented. It consists of the propagation, combination, and satisfaction of three types of constraints: shape-matching constraints, constraints on the degrees of freedom, and nonintersection constraints. Given a high-level description of an assembly and the models of the workpieces, the system determines which parts of the workpieces should be mated and produces a set of homogeneous-coordinate transformations defining the relative position and orientation of each workpiece in the final assembly. This system can be seen as a previous step towards a practical and efficient assembly planner.>

When a set of constraints is imposed on the degrees of freedom between several rigid bodies, finding the configuration or configurations that satisfy all these constraints is a matter of special interest. The problem is not new and has been discussed, not only in kinematics, but also more recently in the design of object-level robot programming languages. In this last domain, several languages have been developed, from different points of view, that are able to partially solve the problem. A more general method is derived than those previously proposed that were based on the symbolic manipulation of chains of matrix products, using the theory of continuous groups.>

Assimilation of a stimulus rhythm by certain nervous structures of vertebrate animals in a conditioning situation has been reported. A lateral-inhibition-type network of plastic pacemaker neurons is proposed to model such behavior. Through simulation, the network exhibits capacity to assimilate and encode in separate groups of neurons two successively presented frequencies. Learning of the second frequency does not disrupt memory of the first one. From the exploration of the effect of varying several factors upon the learning process-related to the connectivity, the intraneuronal functioning, the initial state, and the simulation conditions-it follows that the most influential factors are the proportion of excitatory connections over the total, the ratio between the ranges of the excitatory and the inhibitory connectivity, and the degree of intraneuronal randomness.