IROS 2023 论文列表

The ability to learn new tasks and quickly adapt to different variations or dimensions is an important attribute in agile robotics. In our previous work, we have explored Behavior Trees and Motion Generators (BTMGs) as a robot arm policy representation to facilitate the learning and execution of assembly tasks. The current implementation of the BTMGs for a specific task may not be robust to the changes in the environment and may not generalize well to different variations of tasks. We propose to extend the BTMG policy representation with a module that predicts BTMG parameters for a new task variation. To achieve this, we propose a model that combines a Gaussian process and a weighted support vector machine classifier. This model predicts the performance measure and the feasibility of the predicted policy with BTMG parameters and task variations as inputs. Using the outputs of the model, we then construct a surrogate reward function that is utilized within an optimizer to maximize the performance of a task over BTMG parameters for a fixed task variation. To demonstrate the effectiveness of our proposed approach, we conducted experimental evaluations on push and obstacle avoidance tasks in simulation and with a real KUKA iiwa robot. Furthermore, we compared the performance of our approach with four baseline methods.

Meta-reinforcement learning (meta-RL) is a promising approach that enables the agent to learn new tasks quickly. However, most meta-RL algorithms show poor generalization in multi-task scenarios due to the insufficient task information provided only by rewards. Language-conditioned meta-RL improves the generalization capability by matching language instructions with the agent's behaviors. While both behaviors and language instructions have symmetry, which can speed up human learning of new knowledge. Thus, combining symmetry and language instructions into meta-RL can help improve the algorithm's generalization and learning efficiency. We propose a dual-MDP meta-reinforcement learning method that enables learning new tasks efficiently with symmetrical behav-iors and language instructions. We evaluate our method in mul-tiple challenging manipulation tasks, and experimental results show that our method can greatly improve the generalization and learning efficiency of meta-reinforcement learning. Videos are available at https://tumi6robot.wixsite.com/symmetry/.

In reinforcement learning (RL), sparse rewards can present a significant challenge. Fortunately, expert actions can be utilized to overcome this issue. However, acquiring explicit expert actions can be costly, and expert observations are often more readily available. This paper presents a new approach that uses expert observations for learning in robot manipulation tasks with sparse rewards from pixel observations. Specifically, our technique involves using expert observations as intermediate visual goals for a goal-conditioned RL agent, enabling it to complete a task by successively reaching a series of goals. We demonstrate the efficacy of our method in five challenging block construction tasks in simulation and show that when combined with two state-of-the-art agents, our approach can significantly improve their performance while requiring 4–20 times fewer expert actions during training. Moreover, our method is also superior to a hierarchical baseline.

Learning various motor skills for quadrupedal robots is a challenging problem that requires careful design of task-specific mathematical models or reward descriptions. In this work, we propose to learn a single capable policy using deep reinforcement learning by imitating a large number of reference motions, including walking, turning, pacing, jumping, sitting, and lying. On top of the existing motion imitation framework, we first carefully design the observation space, the action space, and the reward function to improve the scalability of the learning as well as the robustness of the final policy. In addition, we adopt a novel adaptive motion sampling (AMS) method, which maintains a balance between successful and unsuccessful behaviors. This technique allows the learning algorithm to focus on challenging motor skills and avoid catastrophic forgetting. We demonstrate that the learned policy can exhibit diverse behaviors in simulation by successfully tracking both the training dataset and out-of-distribution trajectories. We also validate the importance of the proposed learning formulation and the adaptive motion sampling scheme by conducting experiments.

Cutting objects into desired fragments is challenging for robots due to the spatially unstructured nature of fragments and the complex one-to-many object fragmentation caused by actions. We present a novel approach to model object fragmentation using an attributed stochastic grammar. This grammar abstracts fragment states as node variables and captures causal transitions in object fragmentation through production rules. We devise a probabilistic framework to learn this grammar from human demonstrations. The planning process for object cutting involves inferring an optimal parse tree of desired fragments using the learned grammar, with parse tree productions corresponding to cutting actions. We employ Monte Carlo Tree Search (MCTS) to efficiently approximate the optimal parse tree and generate a sequence of executable cutting actions. The experiments demonstrate the efficacy in planning for object-cutting tasks, both in simulation and on a physical robot. The proposed approach outperforms several baselines by demonstrating superior generalization to novel setups, thanks to the compositionality of the grammar model.

Navigation of terrestrial robots is typically addressed either with localization and mapping (SLAM) followed by classical planning on the dynamically created maps, or by machine learning (ML), often through end-to-end training with reinforcement learning (RL) or imitation learning (IL). Recently, modular designs have achieved promising results, and hybrid algorithms that combine ML with classical planning have been proposed. Existing methods implement these combinations with handcrafted functions, which cannot fully exploit the complementary nature of the policies and the complex regularities between scene structure and planning performance. Our work builds on the hypothesis that the strengths and weaknesses of neural planners and classical planners follow some regularities, which can be learned from training data, in particular from interactions. This is grounded on the assumption that, both, trained planners and the mapping algorithms underlying classical planning are subject to failure cases depending on the semantics of the scene and that this dependence is learnable: for instance, certain areas, objects or scene structures can be reconstructed easier than others. We propose a hierarchical method composed of a high-level planner dynamically switching between a classical and a neural planner. We fully train all neural policies in simulation and evaluate the method in both simulation and real experiments with a LoCoBot robot, showing significant gains in performance, in particular in the real environment. We also qualitatively conjecture on the nature of data regularities exploited by the high-level planner.

Symbolic planning is a powerful technique to solve complex tasks that require long sequences of actions and can equip an intelligent agent with complex behavior. The downside of this approach is the necessity for suitable symbolic representations describing the state of the environment as well as the actions that can change it. Traditionally such representations are carefully hand-designed by experts for distinct problem domains, which limits their transferability to different problems and environment complexities. In this paper, we propose a novel concept to generalize symbolic actions using a given entity hierarchy and observed similar behavior. In a simulated grid-based kitchen environment, we show that type-generalized actions can be learned from few observations and generalize to novel situations. Incorporating an additional on-the-fly generalization mechanism during planning, unseen task combinations, involving longer sequences, novel entities and unexpected environment behavior, can be solved.

In this paper, we present a general learning framework for controlling a quadruped robot that can mimic the behavior of real animals and traverse challenging terrains. Our method consists of two steps: an imitation learning step to learn from motions of real animals, and a terrain adaptation step to enable generalization to unseen terrains. We capture motions from a Labrador on various terrains to facilitate terrain adaptive locomotion. Our experiments demonstrate that our policy can traverse various terrains and produce a natural-looking behavior. We deployed our method on the real quadruped robot \boldsymbol{Max}\boldsymbol{Max} [1] via zero-shot simulation-to-reality transfer, achieving a speed of 1.1 m/s on stairs climbing.

This paper introduces a novel approach for modeling the dynamics of soft robots, utilizing a differentiable filter architecture. The proposed approach enables end-to-end training to learn system dynamics, noise characteristics, and temporal behavior of the robot. A novel spatio-temporal embedding process is discussed to handle observations with varying sensor placements and sampling frequencies. The efficacy of this approach is demonstrated on a tensegrity robot arm by learning end-effector dynamics from demonstrations with complex bending motions. The model is proven to be robust against missing modalities, diverse sensor placement, and varying sampling rates. Additionally, the proposed framework is shown to identify physical interactions with humans during motion. The utilization of a differentiable filter presents a novel solution to the difficulties of modeling soft robot dynamics. Our approach shows substantial improvement in accuracy compared to state-of-the-art filtering methods, with at least a 24% reduction in mean absolute error (MAE) observed. Furthermore, the predicted end-effector positions show an average MAE of 25.77mm from the ground truth, highlighting the advantage of our approach. The code is available at https://github.com/ir-lab/soft_robot_DEnKF.

This study focuses on a robotic powder weighing task used in laboratory automation. In this task, a robot weighs a certain amount of powder with a milligram-level target mass using a dispensing spoon. The complex dynamics of the powder, the variations in the materials being weighed, and the need to balance conservative and aggressive actions are significant challenges in the robotics field. Therefore, learning approaches are critical for this task. However, many learning interactions in real-world environments require substantial efforts to clean the spread powder. To overcome this issue, this study employs a sim-to-real transfer learning approach using a domain randomization (DR) technique. This enables the robot to weigh various powders with a small target mass and alleviates the burden of collecting data in a real-world environment. Herein, we formulated weighing manipulation as a reinforcement learning problem. Besides, we developed a powder weighing simulator and carefully selected the dynamics parameters used for DR to adapt to unseen environments. A recurrent neural network-based policy was adopted considering the balance of conservative and aggressive actions. The sim-to-real zero-shot transfer experiments demonstrated that the robot completed the weighing tasks with an average weighing error of 0.1 - 0.2 mg for different powder materials and target masses (5 - 15 mg). Overall, this approach shows promising results and can be useful for automating laboratory tasks that involve weighing powders.

This study aims to achieve autonomous robotic assembly under uncertain conditions arising from imprecise goal positioning and variations in the angle of the grasped part. Soft robots are suitable for such uncertain and contact-rich environments and are capable of insertion tasks with imprecise goal positions. However, we may also struggle to handle further uncertainty, such as variations in grasping pose. To address the challenge posed by multiple sources of uncertainty, we equipped the soft robot with a tactile sensor. Our key insight is that tactile signal patterns are closely linked to the subtask transitions in an assembly process, specifically from the search to insertion subtasks. We hypothesize soft robots could complete the task by exploring the transition via tactile signals, even in scenarios with imprecise goal positions and grasp misalignment. To this end, we develop an anomaly detection model using a Variational Autoencoder to identify the timing of these transitions. We then employ learning and heuristic-based controllers to navigate the peg tip to the hole and perform the insertion. Our method was validated through real-robot experiments using a soft wrist and a vision-based tactile sensor. The results demonstrate that our method achieves a 100% success rate in scenarios with less uncertain goal pose (\sigma=2\text{mm}\sigma=2\text{mm}) and grasp misalignment (up to 5°) and a 70% success rate in scenarios with uncertain goal pose (\sigma=10\text{mm}\sigma=10\text{mm}) and grasp misalignment (up to 20°). Moreover, our anomaly detection model can generalize to different peg diameters without additional training.

Predicting the future trajectories of surrounding vehicles based on their history trajectories is a critical task in autonomous driving. However, when small crafted perturbations are introduced to those history trajectories, the resulting anomalous (or adversarial) trajectories can significantly mislead the future trajectory prediction module of the ego vehicle, which may result in unsafe planning and even fatal accidents. Therefore, it is of great importance to detect such anomalous trajectories of the surrounding vehicles for system safety, but few works have addressed this issue. In this work, we propose two novel methods for learning effective and efficient representations for online anomaly detection of vehicle trajectories. Different from general time-series anomaly detection, anomalous vehicle trajectory detection deals with much richer contexts on the road and fewer observable patterns on the anomalous trajectories themselves. To address these challenges, our methods exploit contrastive learning techniques and trajectory semantics to capture the patterns underlying the driving scenarios for effective anomaly detection under supervised and unsupervised settings, respectively. We conduct extensive experiments to demonstrate that our supervised method based on contrastive learning and unsupervised method based on reconstruction with semantic latent space can significantly improve the performance of anomalous trajectory detection in their corresponding settings over various baseline methods. We also demonstrate our methods' generalization ability to detect unseen patterns of anomalies.

Deformable robots are notoriously difficult to model or control due to its high-dimensional configuration spaces. Direct trajectory optimization suffers from the curse-of-dimensionality and incurs a high computational cost, while learning-based controller optimization methods are sensitive to hyper-parameter tuning. To overcome these limitations, we hypothesize that high fidelity soft robots can be both simulated and controlled by restricting to low-dimensional spaces. Under such assumption, we propose a two-stage algorithm to identify such simulation- and control-spaces. Our method first identifies the so-called simulation-space that captures the salient deformation modes, to which the robot's governing equation is restricted. We then identify the control-space, to which control signals are restricted. We propose a multi-fidelity Riemannian Bayesian bilevel optimization to identify task-specific control spaces. We show that the dimension of control-space can be less than 10 for a high-DOF soft robot to accomplish walking and swimming tasks, allowing low-dimensional MPC controllers to be applied to soft robots with tractable computational complexity.

The application of reinforcement learning algorithms to robotics has increased over the last decade, especially for the control of robots with non-linear dynamics and a redundant number of degrees of freedom using classic control techniques. Here we study the control of a biomimetic robotic eye with three extraocular muscle pairs as a prime example. Using an actor-critic algorithm, this paper aims to link reinforcement learning to this control problem, and create a framework that will learn the open-loop control of saccadic movements of the robotic eye. The basis for the implemented control is inspired by the primate physiological pulsed control signal, which is generated, integrated and sent to the appropriate muscles to perform the saccade. The metric that evaluates the saccadic output is also inspired by the primate oculomotor system and is used to shape the reward function. This methodology was applied to a simplified 3D physical model of the human eye as a proof of concept. The algorithm managed to learn a saccadic control strategy in 3D. The trajectories obtained, have similar non-linear dynamics as those recorded in humans and their 3D rotational kinematics are constrained by Listing's law.

In this paper, we propose a novel framework for synthesizing a single multimodal control policy capable of generating diverse behaviors (or modes) and emergent inherent transition maneuvers for bipedal locomotion. In our method, we first learn efficient latent encodings for each behavior by training an autoencoder from a dataset of rough reference motions. These latent encodings are used as commands to train a multimodal policy through an adaptive sampling of modes and transitions to ensure consistent performance across different behaviors. We validate the policy's performance in simulation for various distinct locomotion modes such as walking, leaping, jumping on a block, standing idle, and all possible combinations of inter-mode transitions. Finally, we integrate a task-based planner to rapidly generate open-loop mode plans for the trained multimodal policy to solve high-level tasks like reaching a goal position on a challenging terrain. Complex parkour-like motions by smoothly combining the discrete locomotion modes were generated in \sim 3\sim 3 min. to traverse tracks with a gap of width 0.45 m, a plateau of height 0.2 m, and a block of height 0.4 m, which are all significant compared to the dimensions of our mini-biped platform.

The need for opponent modeling and tracking arises in several real-world scenarios, such as professional sports, video game design, and drug-trafficking interdiction. In this work, we present Graph based Adversarial Modeling with Mutual Information (GrAMMI) for modeling the behavior of an adversarial opponent agent. GrAMMI is a novel graph neural network (GNN) based approach that uses mutual information maximization as an auxiliary objective to predict the current and future states of an adversarial opponent with partial observability. To evaluate GrAMMI, we design two large-scale, pursuit-evasion domains inspired by real-world scenarios, where a team of heterogeneous agents is tasked with tracking and interdicting a single adversarial agent, and the adversarial agent must evade detection while achieving its own objectives. With the mutual information formulation, GrAMMI outperforms all baselines in both domains and achieves 31.68% higher log-likelihood on average for future adversarial state predictions across both domains.

This paper presents a study on the use of the Talos humanoid robot for performing assistive sit-to-stand or stand-to-sit tasks. In such tasks, the human exerts a large amount of force (100–200 N) within a very short time (2–8 s), posing significant challenges in terms of human unpredictability and robot stability control. To address these challenges, we propose an approach for finding a spatial reference for the robot, which allows the robot to move according to the force exerted by the human and control its stability during the task. Specifically, we focus on the problem of finding a 1D manifold for the robot, while assuming a simple controller to guide its movement on this manifold. To achieve this, we use a functional representation to parameterize the manifold and solve an optimization problem that takes into account the robot's stability and the unpredictability of human behavior. We demonstrate the effectiveness of our approach through simulations and experiments with the Talos robot, showing robustness and adaptability.

Service robots need language capabilities for communicating with people, and navigation skills for beyond-proximity interaction in the real world. When the robot explores the real world with people side by side, there is the compound problem of human-robot dialog and co-navigation. The human-robot team uses dialog to decide where to go, and their shared spatial awareness affects the dialog state. In this paper, we develop a framework that learns a joint policy for human-robot dialog and co-navigation toward efficiently and accurately completing tour guide and information delivery tasks. We show that our approach outperforms baselines from the literature in task completion rate and execution time, and demonstrate our approach in the real world.

This work addresses human intention identification during physical Human-Robot Interaction (pHRI) tasks to include this information in an assistive controller. To this purpose, human intention is defined as the desired trajectory that the human wants to follow over a finite rolling prediction horizon so that the robot can assist in pursuing it. This work investigates a Recurrent Neural Network (RNN), specifically, Long-Short Term Memory (LSTM) cascaded with a Fully Connected layer. In particular, we propose an iterative training procedure to adapt the model. Such an iterative procedure is powerful in reducing the prediction error. Still, it has the drawback that it is time-consuming and does not generalize to different users or different co-manipulated objects. To overcome this issue, Transfer Learning (TL) adapts the pre-trained model to new trajectories, users, and co-manipulated objects by freezing the LSTM layer and fine-tuning the last FC layer, which makes the procedure faster. Experiments show that the iterative procedure adapts the model and reduces prediction error. Experiments also show that TL adapts to different users and to the co-manipulation of a large object. Finally, to check the utility of adopting the proposed method, we compare the proposed controller enhanced by the intention prediction with the other two standard controllers of pHRI.