Offline Reinforcement Learning (RL) provides a promising avenue for training policies from pre-collected datasets when gathering additional interaction data is infeasible. However, existing offline RL methods often assume stationarity or only consider synthetic perturbations at test time, assumptions that often fail in real-world scenarios characterized by abrupt, time-varying offsets. These offsets can lead to partial observability, causing agents to misperceive their true state and degrade performance. To overcome this challenge, we introduce Forecasting in Non-stationary Offline RL (FORL), a framework that unifies (i) conditional diffusion-based candidate state generation, trained without presupposing any specific pattern of future non-stationarity, and (ii) zero-shot time-series foundation models. FORL targets environments prone to unexpected, potentially non-Markovian offsets, requiring robust agent performance from the onset of each episode. Empirical evaluations on offline RL benchmarks, augmented with real-world time-series data to simulate realistic non-stationarity, demonstrate that FORL consistently improves performance compared to competitive baselines. By integrating zero-shot forecasting with the agent's experience, we aim to bridge the gap between offline RL and the complexities of real-world, non-stationary environments.

Abstract Learning sensorimotor trajectories through flexible neural representations is fundamental for robots as it facilitates the building of motor skills as well as equipping them with the ability to represent the world as predictable temporal events. Recent advances in deep learning led to the development of powerful learning from demonstration (LfD) systems such as Conditional Neural Movement Primitives (CNMPs). CNMPs can robustly represent skills as movement distributions and allow them to be ‘recalled’ by conditioning the movement on a few observation points. In this study, we focus on improving CNMPs to achieve a higher resource economy by adopting a divide-and-conquer approach. We propose a novel neural architecture called Coupled CNMP (C-CNMP), that couples the latent spaces of a pair of CNMPs that splits a given sensorimotor trajectory into segments whose learning is undertaken by smaller sub-networks. Therefore, each sub-network needs to deal with a less complex trajectory making the learning less resource-hungry. With systematic simulations on a controlled trajectory data set, we show that the overhead brought by the coupling introduced in our model is well offset by the resource and performance gain obtained. To be specific, with CNMP model as the baseline, it is shown that the proposed model is able to learn to generate trajectories in the data set with a lower trajectory error measured as the mean absolute difference between the generated trajectory and the ground truth. Importantly, our model can perform well with relatively limited resources, i.e., with less number of neural network parameters compared to the baseline. To show that the findings from the controlled data set well-transfer to robot data, we use robot joint data in an LfD setting and compare the learning performance of the proposed model with the baseline model at equal complexity levels. The simulation experiments show that with also the robot joint data, the proposed model, C-CNMP, learns to generate the joint trajectories with significantly less error than the baseline model. Overall, our study improves the state of the art in sensorimotor trajectory learning and exemplifies how divide-and-conquer approaches can benefit deep learning architectures for resource economy.

Learning from demonstration (LfD) with behavior cloning is attractive for its simplicity; however, compounding errors in long and complex skills can be a hindrance. Considering a target skill as a sequence of motor primitives is helpful in this respect. Then the requirement that a motor primitive ends in a state that allows the successful execution of the subsequent primitive must be met. In this study, we focus on this problem by proposing to learn an explicit correction policy when the expected transition state between primitives is not achieved. The correction policy is learned via behavior cloning by the use of Conditional Neural Motor Primitives (CNMPs) that can generate correction trajectories in a context-dependent way. The advantage of the proposed system over learning the complete task as a single action is shown with a table-top setup in simulation, where an object has to be pushed through a corridor in two steps. Then, the applicability of the proposed method to bi-manual knotting in the real world is shown by equipping an upper-body humanoid robot with the skill of making knots over a bar in 3D space.

Symbolic planning and reasoning are powerful tools for robots tackling complex tasks. However, the need to manually design the symbols restrict their applicability, especially for robots that are expected to act in open-ended environments. Therefore symbol formation and rule extraction should be considered part of robot learning, which, when done properly, will offer scalability, flexibility, and robustness. Towards this goal, we propose a novel general method that finds action-grounded, discrete object and effect categories and builds probabilistic rules over them for non-trivial action planning. Our robot interacts with objects using an initial action repertoire that is assumed to be acquired earlier and observes the effects it can create in the environment. To form action-grounded object, effect, and relational categories, we employ a binary bottleneck layer in a predictive, deep encoderdecoder network that takes the image of the scene and the action applied as input, and generates the resulting effects in the scene in pixel coordinates. After learning, the binary latent vector represents action-driven object categories based on the interaction experience of the robot. To distill the knowledge represented by the neural network into rules useful for symbolic reasoning, a decision tree is trained to reproduce its decoder function. Probabilistic rules are extracted from the decision paths of the tree and are represented in the Probabilistic Planning Domain Definition Language (PPDDL), allowing off-the-shelf planners to operate on the knowledge extracted from the sensorimotor experience of the robot. The deployment of the proposed approach for a simulated robotic manipulator enabled the discovery of discrete representations of object properties such as ‘rollable’ and ‘insertable’. In turn, the use of these representations as symbols allowed the generation of effective plans for achieving goals, such as building towers of the desired height, demonstrating the effectiveness of the approach for multi-step object manipulation. Finally, we demonstrate that the system is not only restricted to the robotics domain by assessing its applicability to the MNIST 8-puzzle domain in which learned symbols allow for the generation of plans that move the empty tile into any given position.

Learning to interact with the environment not only empowers the agent with manipulation capability but also generates information to facilitate building of action understanding and imitation capabilities. This seems to be a strategy adopted by biological systems, in particular primates, as evidenced by the existence of mirror neurons that seem to be involved in multi-modal action understanding. How to benefit from the interaction experience of the robots to enable understanding actions and goals of other agents is still a challenging question. In this study, we propose a novel method, deep modality blending networks (DMBN), that creates a common latent space from multi-modal experience of a robot by blending multi-modal signals with a stochastic weighting mechanism. We show for the first time that deep learning, when combined with a novel modality blending scheme, can facilitate action recognition and produce structures to sustain anatomical and effect-based imitation capabilities. Our proposed system, which is based on conditional neural processes, can be conditioned on any desired sensory/motor value at any time step, and can generate a complete multi-modal trajectory consistent with the desired conditioning in one-shot by querying the network for all the sampled time points in parallel avoiding the accumulation of prediction errors. Based on simulation experiments with an arm-gripper robot and an RGB camera, we showed that DMBN could make accurate predictions about any missing modality (camera or joint angles) given the available ones outperforming recent multimodal variational autoencoder models in terms of long-horizon high-dimensional trajectory predictions. We further showed that given desired images from different perspectives, i.e. images generated by the observation of other robots placed on different sides of the table, our system could generate image and joint angle sequences that correspond to either anatomical or effect-based imitation behavior. To achieve this mirror-like behavior, our system does not perform a pixel-based template matching but rather benefits from and relies on the common latent space constructed by using both joint and image modalities, as shown by additional experiments. Moreover, we showed that mirror learning (in our system) does not only depend on visual experience and cannot be achieved without proprioceptive experience. Our experiments showed that out of ten training scenarios with different initial configurations, the proposed DMBN model could achieve mirror learning in all of the cases where the model that only uses visual information failed in half of them. Overall, the proposed DMBN architecture not only serves as a computational model for sustaining mirror neuron-like capabilities, but also stands as a powerful machine learning architecture for high-dimensional multi-modal temporal data with robust retrieval capabilities operating with partial information in one or multiple modalities.

This paper aims to study the reward-based policy exploration problem in a supervised learning approach and enable robots to form complex movement trajectories in challenging reward settings and search spaces. For this, the experience of the robot, which can be bootstrapped from demonstrated trajectories, is used to train a novel Neural Processes-based deep network that samples from its latent space and generates the required trajectories given desired rewards. Our framework can generate progressively improved trajectories by sampling them from high reward landscapes, increasing the reward gradually. Variational inference is used to create a stochastic latent space to sample varying trajectories in generating a population of trajectories given target rewards. We benefit from Evolutionary Strategies and propose a novel crossover operation, which is applied in the self-organized latent space of the individual policies, allowing blending of the individuals that might address different factors in the reward function. Using a number of tasks that require sequential reaching to multiple points or passing through gaps between objects, we showed that our method provides stable learning progress and significantly higher sample efficiency compared to a number of state-of-the-art robotic reinforcement learning methods. Finally, we show the real-world suitability of our method through real robot execution involving obstacle avoidance.

This study focuses on inferring cost functions of obtained movement data using reward parameter search and pol-icy gradient based Reinforcement Learning (RL). The behavior data for this task is obtained through a series of squat-to-stand movements of human participants under dynamic perturbations. The key parameter searched in the cost function is the weight of total torque used in performing the squat-to-stand action. An approximate model is used to learn squat-to-stand movements via a policy gradient method, namely Proximal Policy Optimization(PPO). A behavioral similarity metric based on Center of Mass(COM) is used to find the most likely weight parameter. The stochasticity in the training result of PPO is dealt with multiple runs, and as a result, a reasonable and a stable Inverse Reinforcement Learning(IRL) algorithm is obtained in terms of performance. The results indicate that for some participants, the reward function parameters of the experts were inferred successfully.

Associative Skill Memories (ASMs) were formulated to encode stereotypical movements along with their stereotypical sensory events to increase the robustness of underlying dynamic movement primitives (DMPs) against noisy perception and perturbations. In ASMs, the stored sensory trajectories, such as the haptic and tactile measurements, are used to compute how much a perturbed movement deviates from the desired one, and to correct the movement if possible. In our work, we extend ASMs: rather than using stored single sensory trajectory instances, our system generates sensory event models and exploits those models to correct the perturbed movements during executions with the aim of generalizing to novel configurations. In particular, measured force and the torque trajectories are modelled using Parametric Hidden Markov Models, and then reproduced by Gaussian Mixture Regression. With Baxter robot, we demonstrate that our proposed force feedback model can be used to correct a trajectory while pushing an object with a mass never experienced before, and which otherwise slips away from the gripper because of noise. In the end, we discuss how far this skill can be generalized using the force model and possible future improvements.

In complex manipulation scenarios (e.g. tasks requiring complex interaction of two hands or in-hand manipulation), generalization is a hard problem. Current methods still either require a substantial amount of (supervised) training data and / or strong assumptions on both the environment and the task. In this paradigm, controllers solving these tasks tend to be complex. We propose a paradigm of maintaining simpler controllers solving the task in a small number of specific situations. In order to generalize to novel situations, the robot transforms the environment from novel situations into a situation where the solution of the task is already known. Our solution to this problem is to play with objects and use previously trained skills (basis skills). These skills can either be used for estimating or for changing the current state of the environment and are organized in skill hierarchies. The approach is evaluated in complex pick-and-place scenarios that involve complex manipulation. We further show that these skills can be learned by autonomous playing.

This work aims for bottom-up and autonomous development of symbolic planning operators from continuous interaction experience of a manipulator robot that explores the environment using its action repertoire. Development of the symbolic knowledge is achieved in two stages. In the first stage, the robot explores the environment by executing actions on single objects, forms effect and object categories, and gains the ability to predict the object/effect categories from the visual properties of the objects by learning the nonlinear and complex relations among them. In the next stage, with further interactions that involve stacking actions on pairs of objects, the system learns logical high-level rules that return a stacking-effect category given the categories of the involved objects and the discrete relations between them. Finally, these categories and rules are encoded in Planning Domain Definition Language (PDDL), enabling symbolic planning. We realized our method by learning the categories and rules in a physics-based simulator. The learned symbols and operators are verified by generating and executing non-trivial symbolic plans on the real robot in a tower building task.

Learning to predict the effects of actions applied to pairs of objects is a difficult task that requires learning complex relations with sparse, incomplete and noisy information. Our Knowledge Propagation approach propagates affordance predictions by exploiting similarities among object properties, action parameters and resulting effects. The knowledge is propagated in a graph where a missing edge, corresponding to an unknown interaction between two objects (nodes), is predicted via the superposition of all paths connecting those objects in the graph. The high complexity of affordance representation is addressed through the use of Maximum Margin Multi-Valued Regression (MMMVR), which scales well to complex problems of multiple layers. With increased diversity and size of object databases and the addition of other parametric combinatory actions, we expect to achieve complex systems that leverage learned structure for subsequent learning, achieving structural bootstrapping over lifelong development and learning. In this paper, we extend MMMVR for learning of paired-object affordances, i.e., for predicting the effects of actions applied to pairs of objects. In our experiments, we evaluated this method on a dataset composed of 83 objects and 83×83 interactions. We compared the prediction performance with standard classifiers that predict the effect category given the object pair's low-level features or single-object affordances. The experiments show that our proposed method achieves significantly higher prediction performance especially when supported with Active Learning.

The direct perception of actions allows a robot to predict the afforded actions of observed objects. In this paper, we present a non-parametric approach to representing the affordance-bearing subparts of objects. This representation forms the basis of a kernel function for computing the similarity between different subparts. Using this kernel function, together with motor primitive actions, the robot can learn the required mappings to perform direct action perception. The proposed approach was successfully implemented on a real robot, which could then quickly learn to generalize grasping and pouring actions to novel objects.

The studies on mirror neurons observed in monkeys indicate that recognition of other's actions activates neural circuits that are also responsible for generating the very same actions in the animal. The mirror neuron hypothesis argues that such an overlap between action generation and recognition can provide a shared worldview among individuals and be a key pillar for communication. Inspired by these findings, this paper extends a learning by demonstration method for online recognition of observed actions. The proposed method is shown to recognize and generate different reaching actions demonstrated by a human on a humanoid robot platform. Experiments show that the proposed method is robust to both occlusions during the observed actions as well as variances in the speed of the observed actions. The results are successfully demonstrated in an interactive game with the iCub humanoid robot platform.

Human infants practice their initial, seemingly random arm movements for transforming them into voluntary reaching and grasping actions. With the developing perceptual abilities, infants further explore their environment using the behavior repertoire they have developed, and learn causality relations in the form of affordances, which they use for goal satisfaction and motor planning. This study proposes and implements a developmental progression on a robotic system mimicking the aforementioned infant development stages: An anthropomorphic robot hand with one basic action of swing-hand and the palmar reflex (i.e. the enclosure of the fingers upon contact) at its disposal, executes swing-hand action targeted to a salient object with different hand speeds. During the executions, it monitors the changes in its sensors, automatically forming behavior primitives such as `grasp', `hit', `carry-object' and `drop' by segmenting and differentiating the initial swing-hand action. The study then focuses on one of these behaviors, namely grasping, and shows how further practice allows the robot to learn affordances of more complex objects, which can be further used to make plans to achieve desired goals using the discovered behavior repertoire.

In this paper, we propose a method that enables a robot to learn not only the existence of affordances provided by objects, but also the behavioral parameters required to actualize them, and the prediction of effects generated on the objects in an unsupervised way. In a previous study, it was shown that through self-interaction and self-observation, analogous to an infant, an anthropomorphic robot can learn object affordances in a completely unsupervised way, and use this knowledge to make plans in its perceptual space. This paper extends the affordances model proposed in that study by using parametric behaviors and including the behavior parameters into affordance learning and goal-oriented plan generation. Furthermore, for handling complex behaviors and complex objects (such as execution of precision grasp on a mug), the perceptual processing is improved by using a combination of local and global features. Finally, a hierarchical clustering algorithm is used to discover the affordances in non-homogenous feature space. In short, object affordances for object manipulation are discovered together with behavior parameters based on the monitored effects.

In this paper, we use the notion of affordances, proposed in cognitive science, as a framework to propose a developmental method that would enable a robot to ground symbolic planning mechanisms in the continuous sensory-motor experiences of a robot. We propose a method that allows a robot to learn the symbolic relations that pertain to its interactions with the world and show that they can be used in planning. Specifically, the robot interacts with the objects in its environment using a pre-coded repertoire of behaviors and records its interactions in a triple that consist of the initial percept of the object, the behavior applied and its effect, defined as the difference between the initial and the final percept. The method allows the robot to learn object affordance relations which can be used to predict the change in the percept of the object when a certain behavior is applied. These relations can then be used to develop plans using forward chaining. The method is implemented and evaluated on a mobile robot system with limited object manipulation capabilities. We have shown that the robot is able to learn the physical affordances of objects from range images and use them to build symbols and relations that can be used in making multi-step predictions about the affordances of objects and achieve complex goals.

"Developmental robotics" proposes that, instead of trying to build a robot that shows intelligence once and for all, what one must do is to build robots that can develop. These robots should be equipped with behaviors that are simple but enough to bootstrap the system. Then, as the robot interacts with its environment, it should display increasingly complex behaviors. In this paper, we propose such a development scheme for a mobile robot. J.J. Gibson's concept of "affordances" and a formalization of this concept provides the basis of this development scheme. We show that an autonomous robot can start with pre-coded primitive behaviors, and as it executes its behaviors randomly in an environment, it can learn the affordance relations between the environment and its behaviors. We then present two ways of using these learned structures, in achieving more complex, intentional behaviors. In the first case, the sequencing of these primitive behaviors are such that new more complex behaviors emerge. In the second case, the robot makes a "blending" of its pre-coded primitive behaviors to create new behaviors.

We are interested in how the concept of affordances can affect our view to autonomous robot control, and how the results obtained from autonomous robotics can be reflected back upon the discussion and studies on the concept of affordances. In this paper, we studied how a mobile robot, equipped with a 3D laser scanner, can learn to perceive the traversability affordance and use it to wander in a room tilled with spheres, cylinders and boxes. The results showed that after learning, the robot can wander around avoiding contact with non-traversable objects (i.e. boxes, upright cylinders, or lying cylinders in certain orientation), but moving over traversable objects (such as spheres, and lying cylinders in a rollable orientation with respect to the robot) rolling them out of its way. We have shown that for each action approximately 1% of the perceptual features were relevant to determine whether it is afforded or not and that these relevant features are positioned in certain regions of the range image. The experiments are conducted both using a physics-based simulator and on a real robot.

In this paper, we studied how a mobile robot equipped with a 3D laser scanner can start from primitive behaviors and learn to use them to achieve goal-directed behaviors. For this purpose, we propose a learning scheme that is based on the concept of "affordances", where the robot first learns about the different kind of effects it can create in the environment and then links these effects with the perception of the initial environment and the executed primitive behavior. It uses these learned relations to create certain effects in the environment and achieve more complex behaviors.