Bias has been shown to be a pervasive problem in machine learning, with severe and unanticipated consequences, for example in the form of algorithm performance disparities across social groups. In this paper, we investigate and characterise how similar issues may arise in Reinforcement Learning (RL) for Human-Robot Interaction (HRI), with the intent of averting the same ramifications. Using an assistive robotics simulation as a case study, we show that RL for HRI can perform differently across models with different waist circumferences. We show this behaviour can arise due to representation bias-unbalanced exposure during training-but also due to inherent task properties that may make assistance difficult depending on physical characteristics. The findings underscore the need to address bias in RL for HRI. We conclude with a discussion of potential practical solutions, their consequences and limitations, and avenues for future research.

While interest in the field of explainable agents increases, it is still an open problem to incorporate a proactive explanation component into a real-time human–agent collaboration. Thus, when collaborating with a human, we want to enable an agent to identify critical moments requiring timely explanations. We differentiate between situations requiring explanations about the agent's decision-making and assistive explanations supporting the user. In order to detect these situations, we analyze eye tracking signals of participants engaging in a collaborative virtual cooking scenario. First, we show how users’ gaze patterns differ between moments of user confusion, the agent making errors, and the user successfully collaborating with the agent. Second, we evaluate different state-of-the-art models on the task of predicting whether the user is confused or the agent makes errors using gaze- and task-related data. An ensemble of MiniRocket classifiers performs best, especially when updating its predictions with high frequency based on input samples capturing time windows of 3 to 5 seconds. We find that gaze is a significant predictor of when and what to explain. Gaze features are crucial to our classifier's accuracy, with task-related features benefiting the classifier to a smaller extent.

Motion planning for deformable object manipulation has been a challenge for a long time in robotics due to its high computational cost. In this work, we propose to mitigate this cost by limiting the number of picking points on a deformable object within the action space and simplifying the dynamics model. We do this first by identifying a minimal geometric model that closely approximates the original model at the goal state; specifically, we implement this general approach for 1-D linear deformable objects (e.g., ropes) using a piece-wise line-fitted model, and for 2-D surface deformable objects (e.g., cloth) using a mesh-simplified model. Then a small number of key particles are extracted as the pickable points in the action space which are sufficient to represent and reach the given goal. Additionally, a simplified dynamics model is constructed based on the simplified geometric model, containing much fewer particles and thus being much faster to simulate than the original dynamics model, albeit with some loss of precision. We further refine this model iteratively by adding more details from the actually achieved final state of the original model until a satisfactory trajectory is generated. Extensive simulation experiments are conducted on a set of representative tasks for ropes and cloth, which show a significant decrease in time cost while achieving similar or better trajectory costs. Finally, we establish a closed-loop system of perception, planning, and control with a real robot for cloth folding, which validates the effectiveness of our proposed method.

Achieving social acceptance is one of the main goals of Social Robotic Navigation. Despite this topic has received increasing interest in recent years, most of the research has focused on driving the robotic agent along obstacle-free trajectories, planning around estimates of future human motion to respect personal distances and optimize navigation. However, social interactions in everyday life are also dictated by norms that do not strictly depend on movement, such as when standing at the end of a queue rather than cutting it. In this paper, we propose a novel method to recognize common social scenarios and modify a traditional planner’s cost function to adapt to them. This solution enables the robot to carry out different social navigation behaviors that would not arise otherwise, maintaining the robustness of traditional navigation. Our approach allows the robot to learn different social norms with a single learned model, rather than having different modules for each task. As a proof of concept, we consider the tasks of queuing and respect interaction spaces of groups of people talking to one another, but the method can be extended to other human activities that do not involve motion.

Modern robots need to adapt to diverse human partners with whom they collaborate. To this end, learning a representation of human capabilities enables the robot to personalize their behaviour to their collaborators across multiple tasks. We propose CApability Modeling from Observations (CAMO), a model-based estimation algorithm, in which human capabilities that parameterize a given model are inferred from observations of the human behaviour on known collaborative tasks. We apply the method to joint limit learning in order to predict future trajectories of a 7-DOF manipulator arm. Furthermore, we show that CAMO can be used as a sub-task assignment routine in a simulated human–robot collaboration scenario, allowing the robot to adapt its task allocation to perform tasks that the person is not able to do.

Planning for deformable object manipulation has been a challenge for a long time in robotics due to its high computational cost. In this work, we propose to reduce this cost by reducing the number of pick points on a deformable object in the action space. We do this by identifying a small number of key particles that are sufficient as pick points to reach a given goal state. We find these key particles through a geometric model simplification process, which finds the minimal geometric model that still enables a good approximation of the original model at the goal state. We present an implementation of this general approach for 1-D linear deformable objects (e.g., ropes) that uses a piece-wise line fitted model, and for 2-D flat deformable objects (e.g., cloth) that uses a mesh simplified model. We conducted simulation experiments on ropes and cloths, which demonstrate the effectiveness of the proposed method. Finally, the planned paths are executed in a real-world setting for two cloth folding tasks.

Service robots are expected to assist users in a constantly growing range of environments and tasks. People may be unique in many ways, and online adaptation of robots is central to personalized assistance. We focus on collaborative tasks in which the human collaborator may not be fully ablebodied, with the aim for the robot to automatically determine the best level of support. We propose a methodology for online adaptation based on Reinforcement Learning and Bayesian inference. As the Reinforcement Learning process continuously adjusts the robot's behavior, the actions that become part of the improved policy are used by the Bayesian inference module as local evidence of human capability, which can be generalized across the state space. The estimated capabilities are then used as pre-conditions to collaborative actions, so that the robot can quickly disable actions that the person seems unable to perform. We demonstrate and validate our approach on two simulated tasks and one real-world collaborative task across a range of motion and sensing capabilities.

Autonomous vehicles use a variety of sensors and machine-learned models to predict the behavior of surrounding road users. Most of the machine-learned models in the literature focus on quantitative error metrics like the root mean square error (RMSE) to learn and report their models’ capabilities. This focus on quantitative error metrics tends to ignore the more important behavioral aspect of the models, raising the question of whether these models really predict human-like behavior. Thus, we propose to analyze the output of machine-learned models much like we would analyze human data in conventional behavioral research. We introduce quantitative metrics to demonstrate presence of three different behavioral phenomena in a naturalistic highway driving dataset: 1) The kinematics-dependence of who passes a merging point first 2) Lane change by an on-highway vehicle to accommodate an on-ramp vehicle 3) Lane changes by vehicles on the highway to avoid lead vehicle conflicts. Then, we analyze the behavior of three machine-learned models using the same metrics. Even though the models’ RMSE value differed, all the models captured the kinematic-dependent merging behavior but struggled at varying degrees to capture the more nuanced courtesy lane change and highway lane change behavior. Additionally, the collision aversion analysis during lane changes showed that the models struggled to capture the physical aspect of human driving: leaving adequate gap between the vehicles. Thus, our analysis highlighted the inadequacy of simple quantitative metrics and the need to take a broader behavioral perspective when analyzing machine-learned models of human driving predictions.

As human–agent collaborations become more prevalent, it is increasingly important for an agent to be able to adapt to their collaborator and explain their own behavior. In order to do so, they need to be able to identify critical states during the interaction that call for proactive clarifications or behavioral adaptations. In this paper, we explore whether the agent could infer such states from the human’s eye gaze for which we compare gaze patterns across different situations in a collaborative task. Our findings show that the human’s gaze patterns significantly differ between times at which the user is confused about the task, times at which the agent makes an error, and times of normal workflow. During errors the amount of gaze towards the agent increases, while during confusion the amount towards the environment increases. We conclude that these signals could tell the agent what and when to explain.

We address the manipulation task of retrieving a target object from a cluttered shelf. When the target object is hidden, the robot must search through the clutter for retrieving it. Solving this task requires reasoning over the likely locations of the target object. It also requires physics reasoning over multi-object interactions and future occlusions. In this work, we present a data-driven hybrid planner for generating occlusion-aware actions in closed-loop. The hybrid planner explores likely locations of the occluded target object as predicted by a learned distribution from the observation stream. The search is guided by a heuristic trained with reinforcement learning to act on observations with occlusions. We evaluate our approach in different simulation and real-world settings (video available on https://youtu.be/dY7YQ3LUVQg). The results validate that our approach can search and retrieve a target object in near real time in the real world while only being trained in simulation.

Humans, in comparison to robots, are remarkably adept at reaching for objects in cluttered environments. The best existing robot planners are based on random sampling of configuration space- which becomes excessively high-dimensional with large number of objects. Consequently, most planners often fail to efficiently find object manipulation plans in such environments. We addressed this problem by identifying high-level manipulation plans in humans, and transferring these skills to robot planners. We used virtual reality to capture human participants reaching for a target object on a tabletop cluttered with obstacles. From this, we devised a qualitative representation of the task space to abstract the decision making, irrespective of the number of obstacles. Based on this representation, human demonstrations were segmented and used to train decision classifiers. Using these classifiers, our planner produced a list of waypoints in task space. These waypoints provided a high-level plan, which could be transferred to an arbitrary robot model and used to initialise a local trajectory optimiser. We evaluated this approach through testing on unseen human VR data, a physics-based robot simulation, and a real robot (dataset and code are publicly available1). We found that the human-like planner outperformed a state-of-the-art standard trajectory optimisation algorithm, and was able to generate effective strategies for rapid planning- irrespective of the number of obstacles in the environment.

In Meta-Reinforcement Learning (meta-RL) an agent is trained on a set of tasks to prepare for and learn faster in new, unseen, but related tasks. The training tasks are usually hand-crafted to be representative of the expected distribution of target tasks and hence all used in training. We show that given a set of training tasks, learning can be both faster and more effective (leading to better performance in the target tasks), if the training tasks are appropriately selected. We propose a task selection algorithm based on information theory, which optimizes the set of tasks used for training in meta-RL, irrespectively of how they are generated. The algorithm establishes which training tasks are both sufficiently relevant for the target tasks, and different enough from one another. We reproduce different meta-RL experiments from the literature and show that our task selection algorithm improves the final performance in all of them.

Reinforcement learning (RL) is a popular paradigm for addressing sequential decision tasks in which the agent has only limited environmental feedback. Despite many advances over the past three decades, learning in many domains still requires a large amount of interaction with the environment, which can be prohibitively expensive in realistic scenarios. To address this problem, transfer learning has been applied to reinforcement learning such that experience gained in one task can be leveraged when starting to learn the next, harder task. More recently, several lines of research have explored how tasks, or data samples themselves, can be sequenced into a curriculum for the purpose of learning a problem that may otherwise be too difficult to learn from scratch. In this article, we present a framework for curriculum learning (CL) in reinforcement learning, and use it to survey and classify existing CL methods in terms of their assumptions, capabilities, and goals. Finally, we use our framework to find open problems and suggest directions for future RL curriculum learning research.

Curriculum learning has been successfully used in reinforcement learning to accelerate the learning process, through knowledge transfer between tasks of increasing complexity. Critical tasks, in which suboptimal exploratory actions must be minimized, can benefit from curriculum learning, and its ability to shape exploration through transfer. We propose a task sequencing algorithm maximizing the cumulative return, that is, the return obtained by the agent across all the learning episodes. By maximizing the cumulative return, the agent not only aims at achieving high rewards as fast as possible, but also at doing so while limiting suboptimal actions. We experimentally compare our task sequencing algorithm to several popular metaheuristic algorithms for combinatorial optimization, and show that it achieves significantly better performance on the problem of cumulative return maximization. Furthermore, we validate our algorithm on a critical task, optimizing a home controller for a micro energy grid.

Physics-based manipulation in clutter involves complex interaction between multiple objects. In this paper, we consider the problem of learning, from interaction in a physics simulator, manipulation skills to solve this multi-step sequential decision making problem in the real world. Our approach has two key properties: (i) the ability to generalize and transfer manipulation skills (over the type, shape, and number of objects in the scene) using an abstract image-based representation that enables a neural network to learn useful features; and (ii) the ability to perform look-ahead planning in the image space using a physics simulator, which is essential for such multi-step problems. We show, in sets of simulated and real-world experiments (video available on https://youtu.be/EmkUQfyvwkY), that by learning to evaluate actions in an abstract image-based representation of the real world, the robot can generalize and adapt to the object shapes in challenging real-world environments.

In recent years, research has shown that transfer learning methods can be leveraged to construct curricula that sequence a series of simpler tasks such that performance on a final target task is improved. A major limitation of existing approaches is that such curricula are handcrafted by humans that are typically domain experts. To address this limitation, we introduce a method to generate a curriculum based on task descriptors and a novel metric of transfer potential. Our method automatically generates a curriculum as a directed acyclic graph (as opposed to a linear sequence as done in existing work). Experiments in both discrete and continuous domains show that our method produces curricula that improve the agent's learning performance when compared to the baseline condition of learning on the target task from scratch.

We investigate methods to improve fault-tolerance of Autonomous Underwater Vehicles (AUVs) to increase their reliability and persistent autonomy. We propose a learning-based approach that is able to discover new control policies to overcome thruster failures as they happen. The proposed approach is a model-based direct policy search that learns on an on-board simulated model of the AUV. The model is adapted to a new condition when a fault is detected and isolated. Since the approach generates an optimal trajectory, the learned fault-tolerant policy is able to navigate the AUV towards a specified target with minimum cost. Finally, the learned policy is executed on the real robot in a closed-loop using the state feedback of the AUV. Unlike most existing methods which rely on the redundancy of thrusters, our approach is also applicable when the AUV becomes under-actuated in the presence of a fault. To validate the feasibility and efficiency of the presented approach, we evaluate it with three learning algorithms and three policy representations with increasing complexity. The proposed method is tested on a real AUV, Girona500.

We describe the design and implementation of an on-line identification scheme for Autonomous Underwater Vehicles (AUVs). The proposed method estimates the dynamic parameters of the vehicle based on a global derivative-free optimization algorithm. It is not sensitive to initial conditions, unlike other on-line identification schemes, and does not depend on the differentiability of the model with respect to the parameters. The identification scheme consists of three distinct modules: a) System Excitation, b) Metric Calculator and c) Optimization Algorithm. The System Excitation module sends excitation inputs to the vehicle. The Optimization Algorithm module calculates a candidate parameter vector, which is fed to the Metric Calculator module. The Metric Calculator module evaluates the candidate parameter vector, using a metric based on the residual of the actual and the predicted commands. The predicted commands are calculated utilizing the candidate parameter vector and the vehicle state vector, which is available via a complete navigation module. Then, the metric is directly fed back to the Optimization Algorithm module, and it is used to correct the estimated parameter vector. The procedure continues iteratively until the convergence properties are met. The proposed method is generic, demonstrates quick convergence and does not require a linear formulation of the model with respect to the parameter vector. The applicability and performance of the proposed algorithm is experimentally verified using the AUV Girona 500.

The execution of tasks for a robotic agent embedded in a dynamic environment brings about several challenges, due to unpredictable (or unobservable) events, and to inaccurate perception. Moreover, the agent can perform multiple tasks and each task can be achieved by applying different plans, therefore the decision about which strategy is the most convenient, given the current situation of the world, is important for assessing an intelligent overall behavior of the agent. This paper tackles the problem of on-line execution monitoring in a novel way with respect to previous work, since: (1) it considers uncertainty in the duration of actions with a probabilistic model of action duration; (2) it evaluates the cost of each possible plan at run-time in terms of probability of successful termination within a desired expected time. The approach has been evaluated both in a robotic soccer and a surveillance scenario.