Reinforcement learning is inherently unsafe for use in physical systems, as learning by trial-and-error can cause harm to the environment or the robot itself. One way to avoid unpredictable exploration is to add constraints in the action space to restrict the robot behavior. In this paper, we proposea null space based framework of integrating reinforcement learning methods in constrained continuous action spaces. We leverage a hierarchical control framework to decompose target robotic skills into higher ranked tasks (e. g., joint limits and obstacle avoidance) and lower ranked reinforcement learning task. Safe exploration is guaranteed by only learning policies in the null space of higher prioritized constraints. Meanwhile multiple constraint phases for different operational spaces are constructed to guide the robot exploration. Also, we add penalty loss for violating higher ranked constraints to accelerate the learning procedure. We have evaluated our method on different redundant robotic tasks in simulation and show that our null space based reinforcement learning method can explore and learn safely and efficiently.

Contact-rich manipulation tasks remain a hard problem in robotics that requires interaction with unstructured environments. Reinforcement Learning (RL) is one potential solution to such problems, as it has been successfully demonstrated on complex continuous control tasks. Nevertheless, current state-of-the-art methods require policy training in simulation to prevent undesired behavior and later domain transfer even for simple skills involving contact. In this paper, we address the problem of learning contact-rich manipulation policies by extending an existing skill-based RL framework with a variable impedance action space. Our method leverages a small set of suboptimal demonstration trajectories and learns from both position, but also crucially impedance-space information. We evaluate our method on a number of peg-in-hole task variants with a Franka Panda arm and demonstrate that learning variable impedance actions for RL in Cartesian space can be deployed directly on the real robot, without resorting to learning in simulation.

In manufacturing, assembly tasks have been a challenge for learning algorithms due to variant dynamics of different environments. Reinforcement learning (RL) is a promising framework to automatically learn these tasks, yet it is still not easy to apply a learned policy or skill, that is the ability of solving a task, to a similar environment even if the deployment conditions are only slightly different. In this letter, we address the challenge of transferring knowledge within a family of similar tasks by leveraging multiple skill priors. We propose to learn prior distribution over the specific skill required to accomplish each task and compose the family of skill priors to guide learning the policy for a new task by comparing the similarity between the target task and the prior ones. Our method learns a latent action space representing the skill embedding from demonstrated trajectories for each prior task. We have evaluated our method on a task in simulation and a set of peg-in-hole insertion tasks and demonstrate better generalization to new tasks that have never been encountered during training. Our Multi-Prior Regularized RL (MPR-RL) method is deployed directly on a real world Franka Panda arm, requiring only a set of demonstrated trajectories from similar, but crucially not identical, problem instances.

Reinforcement learning (RL) has shown impressive results on a variety of robot tasks, but it requires a large amount of data for learning a single RL policy. However, in manufacturing there is a wide demand of reusing skills from different robots and it is hard to transfer the learned policy to different hardware due to diverse robot body morphology, kinematics, and dynamics. In this paper, we address the problem of transferring policies between different robot platforms. We learn a set of skills on each specific robot and represent them in a latent space. We propose to transfer the skills between different robots by mapping latent action spaces through a cycle generative network in a supervised learning manner. We extend the policy model learned on one robot with a pre-trained generative network to enable the robot to learn from the skill of another robot. We evaluate our method on several simulated experiments and demonstrate that our Learn from Robot (LfR) method accelerates new skill learning.