Human brain and behavior provide a rich venue that can inspire novel control and learning methods for robotics. In an attempt to exemplify such a development by inspiring how humans acquire knowledge and transfer skills among tasks, we introduce a novel multi-task reinforcement learning framework named Episodic Return Progress with Bidirectional Progressive Neural Networks (ERP-BPNN). The proposed ERP-BPNN model (1) learns in a human-like interleaved manner by (2) autonomous task switching based on a novel intrinsic motivation signal and, in contrast to existing methods, (3) allows bidirectional skill transfer among tasks. ERP-BPNN is a general architecture applicable to several multi-task learning settings; in this paper, we present the details of its neural architecture and show its ability to enable effective learning and skill transfer among morphologically different robots in a reaching task. The developed Bidirectional Progressive Neural Network (BPNN) architecture enables bidirectional skill transfer without requiring incremental training and seamlessly integrates with online task arbitration. The task arbitration mechanism developed is based on soft Episodic Return progress (ERP), a novel intrinsic motivation (IM) signal. To evaluate our method, we use quantifiable robotics metrics such as 'expected distance to goal' and 'path straightness' in addition to the usual reward-based measure of episodic return common in reinforcement learning. With simulation experiments, we show that ERP-BPNN achieves faster cumulative convergence and improves performance in all metrics considered among morphologically different robots compared to the baselines.
Discovering the symbols and rules that can be used in long-horizon planning from a robot's unsupervised exploration of its environment and continuous sensorimotor experience is a challenging task. The previous studies proposed learning symbols from single or paired object interactions and planning with these symbols. In this work, we propose a system that learns rules with discovered object and relational symbols that encode an arbitrary number of objects and the relations between them, converts those rules to Planning Domain Description Language (PDDL), and generates plans that involve affordances of the arbitrary number of objects to achieve tasks. We validated our system with box-shaped objects in different sizes and showed that the system can develop a symbolic knowledge of pick-up, carry, and place operations, taking into account object compounds in different configurations, such as boxes would be carried together with a larger box that they are placed on. We also compared our method with the state-of-the-art methods and showed that planning with the operators defined over relational symbols gives better planning performance compared to the baselines.
We observe a large variety of robots in terms of their bodies, sensors, and actuators. Given the commonalities in the skill sets, teaching each skill to each different robot independently is inefficient and not scalable when the large variety in the robotic landscape is considered. If we can learn the correspondences between the sensorimotor spaces of different robots, we can expect a skill that is learned in one robot can be more directly and easily transferred to the other robots. In this paper, we propose a method to learn correspondences between robots that have significant differences in their morphologies: a fixed-based manipulator robot with joint control and a differential drive mobile robot. For this, both robots are first given demonstrations that achieve the same tasks. A common latent representation is formed while learning the corresponding policies. After this initial learning stage, the observation of a new task execution by one robot becomes sufficient to generate a latent space representation pertaining to the other robot to achieve the same task. We verified our system in a set of experiments where the correspondence between two simulated robots is learned (1) when the robots need to follow the same paths to achieve the same task, (2) when the robots need to follow different trajectories to achieve the same task, and (3) when complexities of the required sensorimotor trajectories are different for the robots considered. We also provide a proof-of-the-concept realization of correspondence learning between a real manipulator robot and a simulated mobile robot.
Exploratoration and self-observation are key mechanisms of infant sensorimotor development. These processes are further guided by parental scaffolding accelerating skill and knowledge acquisition. In developmental robotics, this approach has been adopted often by having a human acting as the source of scaffolding. In this study, we investigate whether Large Language Models (LLMs) can act as a scaffolding agent for a robotic system that aims to learn to predict the effects of its actions. To this end, an object manipulation setup is considered where one object can be picked and placed on top of or in the vicinity of another object. The adopted LLM is asked to guide the action selection process through algorithmically generated state descriptions and action selection alternatives in natural language. The simulation experiments that include cubes in this setup show that LLM-guided (GPT3.5-guided) learning yields significantly faster discovery of novel structures compared to random exploration. However, we observed that GPT3.5 fails to effectively guide the robot in generating structures with different affordances such as cubes and spheres. Overall, we conclude that even without fine-tuning, LLMs may serve as a moderate scaffolding agent for improving robot learning, however, they still lack affordance understanding which limits the applicability of the current LLMs in robotic scaffolding tasks.
In this paper, we propose and realize a new deep learning architecture for discovering symbolic representations for objects and their relations based on the self-supervised continuous interaction of a manipulator robot with multiple objects on a tabletop environment. The key feature of the model is that it can handle a changing number number of objects naturally and map the object-object relations into symbolic domain explicitly. In the model, we employ a self-attention layer that computes discrete attention weights from object features, which are treated as relational symbols between objects. These relational symbols are then used to aggregate the learned object symbols and predict the effects of executed actions on each object. The result is a pipeline that allows the formation of object symbols and relational symbols from a dataset of object features, actions, and effects in an end-to-end manner. We compare the performance of our proposed architecture with state-of-the-art symbol discovery methods in a simulated tabletop environment where the robot needs to discover symbols related to the relative positions of objects to predict the observed effect successfully. Our experiments show that the proposed architecture performs better than other baselines in effect prediction while forming not only object symbols but also relational symbols. Furthermore, we analyze the learned symbols and relational patterns between objects to learn about how the model interprets the environment. Our analysis shows that the learned symbols relate to the relative positions of objects, object types, and their horizontal alignment on the table, which reflect the regularities in the environment.
Offline Reinforcement Learning (RL) methods leverage previous experiences to learn better policies than the behavior policy used for experience collection. In contrast to behavior cloning, which assumes the data is collected from expert demonstrations, offline RL can work with non-expert data and multimodal behavior policies. However, offline RL algorithms face challenges in handling distribution shifts and effectively representing policies due to the lack of online interaction during training. Prior work on offline RL uses conditional diffusion models to obtain expressive policies to represent multimodal behavior in the dataset. Nevertheless, they are not tailored toward alleviating the out-of-distribution state generalization. We introduce a novel method incorporating state reconstruction feature learning in the recent class of diffusion policies to address the out-of-distribution generalization problem. State reconstruction loss promotes more descriptive representation learning of states to alleviate the distribution shift incurred by the out-of-distribution states. We design a 2D Multimodal Contextual Bandit environment to demonstrate and evaluate our proposed model. We assess the performance of our model not only in this new environment but also on several D4RL benchmark tasks, achieving state-of-the-art results.
Learning from demonstration (LfD) provides a convenient means to equip robots with dexterous skills when demonstration can be obtained in robot intrinsic coordinates. However, the problem of compounding errors in long and complex skills reduces its wide deployment. Since most such complex skills are composed of smaller movements that are combined, considering the target skill as a sequence of compact motor primitives seems reasonable. Here the problem that needs to be tackled is to ensure that a motor primitive ends in a state that allows the successful execution of the subsequent primitive. 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 itself learned via behavior cloning by the use of a state-of-the-art movement primitive learning architecture, Conditional Neural Motor Primitives (CNMPs). The learned correction policy is then able to produce diverse movement 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. The experiments show that the robot can perform successful knotting even when the faced correction cases are not part of the human demonstration set.
In this paper, we propose a concept learning architecture that enables a robot to build symbols through self-exploration by interacting with a varying number of objects. Our aim is to allow a robot to learn concepts without constraints, such as a fixed number of interacted objects or pre-defined symbolic structures. As such, the sought architecture should be able to build symbols for objects such as single objects that can be grasped, object stacks that cannot be grasped together, or other composite dynamic structures. Towards this end, we propose a novel architecture, a self-attentive predictive encoder-decoder network with binary activation layers. We show the validity of the proposed network through a robotic manipulation setup involving a varying number of rigid objects. The continuous sensorimotor experience of the robot is used by the proposed network to form effect predictors and symbolic structures that describe the interaction of the robot in a discrete way. We showed that the robot acquired reasoning capabilities to encode interaction dynamics of a varying number of objects in different configurations using the discovered symbols. For example, the robot could reason that (possible multiple numbers of) objects on top of another object would move together if the object below is moved by the robot. We also showed that the discovered symbols can be used for planning to reach goals by training a higher-level neural network that makes pure symbolic reasoning.
Abstraction is an important aspect of intelligence which enables agents to construct robust representations for effective decision making. In the last decade, deep networks are proven to be effective due to their ability to form increasingly complex abstractions. However, these abstractions are distributed over many neurons, making the re-use of a learned skill costly. Previous work either enforced formation of abstractions creating a designer bias, or used a large number of neural units without investigating how to obtain high-level features that may more effectively capture the source task. For avoiding designer bias and unsparing resource use, we propose to exploit neural response dynamics to form compact representations to use in skill transfer. For this, we consider two competing methods based on (1) maximum information compression principle and (2) the notion that abstract events tend to generate slowly changing signals, and apply them to the neural signals generated during task execution. To be concrete, in our simulation experiments, we either apply principal component analysis (PCA) or slow feature analysis (SFA) on the signals collected from the last hidden layer of a deep network while it performs a source task, and use these features for skill transfer in a new target task. We compare the generalization performance of these alternatives with the baselines of skill transfer with full layer output and no-transfer settings. Our results show that SFA units are the most successful for skill transfer. SFA as well as PCA, incur less resources compared to usual skill transfer, whereby many units formed show a localized response reflecting end-effector-obstacle-goal relations. Finally, SFA units with lowest eigenvalues resembles symbolic representations that highly correlate with high-level features such as joint angles which might be thought of precursors for fully symbolic systems.
Autonomous discovery of discrete symbols and rules from continuous interaction experience is a crucial building block of robot AI, but remains a challenging problem. Solving it will overcome the limitations in scalability, flexibility, and robustness of manually-designed symbols and rules, and will constitute a substantial advance towards autonomous robots that can learn and reason at abstract levels in open-ended environments. Towards this goal, we propose a novel and general method that finds action-grounded, discrete object and effect categories and builds probabilistic rules over them that can be used in complex action planning. Our robot interacts with single and multiple objects using a given action repertoire and observes the effects created in the environment. In order to form action-grounded object, effect, and relational categories, we employ a binarized bottleneck layer of a predictive, deep encoder-decoder network that takes as input the image of the scene and the action applied, and generates the resulting object displacements in the scene (action effects) in pixel coordinates. The binary latent vector represents a learned, action-driven categorization of objects. To distill the knowledge represented by the neural network into rules useful for symbolic reasoning, we train a decision tree to reproduce its decoder function. From its branches we extract probabilistic rules and represent them in PPDDL, allowing off-the-shelf planners to operate on the robot's sensorimotor experience. Our system is verified in a physics-based 3d simulation environment where a robot arm-hand system learned symbols that can be interpreted as 'rollable', 'insertable', 'larger-than' from its push and stack actions; and generated effective plans to achieve goals such as building towers from given cubes, balls, and cups using off-the-shelf probabilistic planners.