Simulation to Real-World Transfer allows affordable and fast training of learning-based robots for manipulation tasks using Deep Reinforcement Learning methods. Currently, Sim2Real uses Asymmetric Actor-Critic approaches to reduce the rich idealized features in simulation to the accessible ones in the real world. However, the feature reduction from the simulation to the real world is conducted through an empirically defined one-step curtail. Small feature reduction does not sufficiently remove the actor's features, which may still cause difficulty setting up the physical system, while large feature reduction may cause difficulty and inefficiency in training. To address this issue, we proposed Curriculum-based Sensing Reduction to enable the actor to start with the same rich feature space as the critic and then get rid of the hard-to-extract features step-by-step for higher training performance and better adaptation for real-world feature space. The reduced features are replaced with random signals from a Deep Random Generator to remove the dependency between the output and the removed features and avoid creating new dependencies. The methods are evaluated on the Allegro robot hand in a real-world in-hand manipulation task. The results show that our methods have faster training and higher task performance than baselines and can solve real-world tasks when selected tactile features are reduced.
Deep Reinforcement Learning has shown its capability to solve the high degrees of freedom in control and the complex interaction with the object in the multi-finger dexterous in-hand manipulation tasks. Current DRL approaches prefer sparse rewards to dense rewards for the ease of training but lack behavior constraints during the manipulation process, leading to aggressive and unstable policies that are insufficient for safety-critical in-hand manipulation tasks. Dense rewards can regulate the policy to learn stable manipulation behaviors with continuous reward constraints but are hard to empirically define and slow to converge optimally. This work proposes the Finger-specific Multi-agent Shadow Reward (FMSR) method to determine the stable manipulation constraints in the form of dense reward based on the state-action occupancy measure, a general utility of DRL that is approximated during the learning process. Information Sharing (IS) across neighboring agents enables consensus training to accelerate the convergence. The methods are evaluated in two in-hand manipulation tasks on the Shadow Hand. The results show FMSR+IS converges faster in training, achieving a higher task success rate and better manipulation stability than conventional dense reward. The comparison indicates FMSR+IS achieves a comparable success rate even with the behavior constraint but much better manipulation stability than the policy trained with a sparse reward.
Although data-driven motion mapping methods are promising to allow intuitive robot control and teleoperation that generate human-like robot movement, they normally require tedious pair-wise training for each specific human and robot pair. This paper proposes a transferability-based mapping scheme to allow new robot and human input systems to leverage the mapping of existing trained pairs to form a mapping transfer chain, which will reduce the number of new pair-specific mappings that need to be generated. The first part of the mapping schematic is the development of a Synergy Mapping via Dual-Autoencoder (SyDa) method. This method uses the latent features from two autoencoders to extract the common synergy of the two agents. Secondly, a transferability metric is created that approximates how well the mapping between a pair of agents will perform compared to another pair before creating the motion mapping models. Thus, it can guide the formation of an optimal mapping chain for the new human-robot pair. Experiments with human subjects and a Pepper robot demonstrated 1) The SyDa method improves the accuracy and generalizability of the pair mappings, 2) the SyDa method allows for bidirectional mapping that does not prioritize the direction of mapping motion, and 3) the transferability metric measures how compatible two agents are for accurate teleoperation. The combination of the SyDa method and transferability metric creates generalizable and accurate mapping need to create the transfer mapping chain.
In-hand manipulation is challenging for a multi-finger robotic hand due to its high degrees of freedom and the complex interaction with the object. To enable in-hand manipulation, existing deep reinforcement learning based approaches mainly focus on training a single robot-structure-specific policy through the centralized learning mechanism, lacking adaptability to changes like robot malfunction. To solve this limitation, this work treats each finger as an individual agent and trains multiple agents to control their assigned fingers to complete the in-hand manipulation task cooperatively. We propose the Multi-Agent Global-Observation Critic and Local-Observation Actor (MAGCLA) method, where the critic can observe all agents' actions globally, and the actor only locally observes its neighbors' actions. Besides, conventional individual experience replay may cause unstable cooperation due to the asynchronous performance increment of each agent, which is critical for in-hand manipulation tasks. To solve this issue, we propose the Synchronized Hindsight Experience Replay (SHER) method to synchronize and efficiently reuse the replayed experience across all agents. The methods are evaluated in two in-hand manipulation tasks on the Shadow dexterous hand. The results show that SHER helps MAGCLA achieve comparable learning efficiency to a single policy, and the MAGCLA approach is more generalizable in different tasks. The trained policies have higher adaptability in the robot malfunction test compared to the baseline multi-agent and single-agent approaches.
Autonomous grasping is challenging due to the high computational cost caused by multi-fingered robotic hands and their interactions with objects. Various analytical methods have been developed yet their high computational cost limits the adoption in real-world applications. Learning-based grasping can afford real-time motion planning thanks to its high computational efficiency. However, it needs to explore large search spaces during its learning process. The search space causes low learning efficiency, which has been the main barrier to its practical adoption. In this work, we develop a novel Physics-Guided Deep Reinforcement Learning with a Hierarchical Reward Mechanism, which combines the benefits of both analytical methods and learning-based methods for autonomous grasping. Different from conventional observation-based grasp learning, physics-informed metrics are utilized to convey correlations between features associated with hand structures and objects to improve learning efficiency and learning outcomes. Further, a hierarchical reward mechanism is developed to enable the robot to learn the grasping task in a prioritized way. It is validated in a grasping task with a MICO robot arm in simulation and physical experiments. The results show that our method outperformed the baseline in task performance by 48% and learning efficiency by 40%.
Dexterous manipulation tasks usually have multiple objectives, and the priorities of these objectives may vary at different phases of a manipulation task. Varying priority makes a robot hardly or even failed to learn an optimal policy with a deep reinforcement learning (DRL) method. To solve this problem, we develop a novel Adaptive Hierarchical Reward Mechanism (AHRM) to guide the DRL agent to learn manipulation tasks with multiple prioritized objectives. The AHRM can determine the objective priorities during the learning process and update the reward hierarchy to adapt to the changing objective priorities at different phases. The proposed method is validated in a multi-objective manipulation task with a JACO robot arm in which the robot needs to manipulate a target with obstacles surrounded. The simulation and physical experiment results show that the proposed method improved robot learning in task performance and learning efficiency.
In human-robot cooperation, the robot cooperates with the human to accomplish the task together. Existing approaches assume the human has a specific goal during the cooperation, and the robot infers and acts toward it. However, in real-world environments, a human usually only has a general goal (e.g., general direction or area in motion planning) at the beginning of the cooperation which needs to be clarified to a specific goal (e.g., an exact position) during cooperation. The specification process is interactive and dynamic, which depends on the environment and the behavior of the partners. The robot that does not consider the goal specification process may cause frustration to the human partner, elongate the time to come to an agreement, and compromise or fail team performance. We present Evolutionary Value Learning (EVL) approach which uses a State-based Multivariate Bayesian Inference method to model the dynamics of goal specification process in HRC, and an Evolutionary Value Updating method to actively enhance the process of goal specification and cooperation formation. This enables the robot to simultaneously help the human to specify the goal and learn a cooperative policy in a Reinforcement Learning manner. In experiments with real human subjects, the robot equipped with EVL outperforms existing methods with faster goal specification processes and better team performance.
Increasing the autonomy level of a robot hand to accomplish remote object manipulation tasks faster and easier is a new and promising topic in teleoperation. Such semi-autonomous telemanipulation, however, is very challenging due to the physical discrepancy between the human hand and the robot hand, along with the fine motion constraints required for the manipulation task. To overcome these challenges, the robot needs to learn how to assist the human operator in a preferred/intuitive way, which must provide effective assistance that the operator needs yet still accommodate human inputs, so the operator feels in control of the system (i.e., not counter-intuitive to the operator). Toward this goal, we develop novel data-driven approaches to stably learn what assistance is preferred from high data variance caused by the ambiguous nature of human operators. To avoid an extensive robot-specific training process, methods to transfer this assistance knowledge between different robot hands are discussed. Experiments were conducted to telemanipulate a cup for three principal tasks: usage, move, and handover by remotely controlling a 3-finger gripper and 2-finger gripper. Results demonstrated that the proposed model effectively learned the knowledge of preferred assistance, and knowledge transfer between robots allows this semi-autonomous telemanipulation strategy to be scaled up with less training efforts.
The deep reinforcement learning method for human-robot cooperation (HRC) is promising for its high performance when robots are learning complex tasks. However, the applicability of such an approach in a real-world context is limited due to long training time, additional training difficulty caused by inconsistent human performance and the inherent instability of policy exploration. With this approach, the robot has two dynamics to learn: how to accomplish the given physical task and how to cooperate with the human partner. Furthermore, the dynamics of the task and human partner are usually coupled, which means the observable outcomes and behaviors are coupled. It is hard for the robot to efficiently learn from coupled observations. In this paper, we hypothesize that the robot needs to learn the task separately from learning the behavior of the human partner to improve learning efficiency and outcomes. This leads to a fundamental question: Should the robot learn the task first or learn the human behavior first (Fig. 1)? We develop a novel hierarchical rewards mechanism with a task decomposition method that enables the robot to efficiently learn a complex hierarchical dynamic task and human behavior for better HRC. The algorithm is validated in a hierarchical control task in a simulated environment with human subject experiments, and we are able to answer the question by analyzing the collected experiment results.