Failure Prevention in Bimanual Robots using Deep Deterministic Policy Gradient

One of the key challenges in learning robot object interaction behavior is to obtain an optimal policy that accomplishes the goal safely. In reinforcement learning, representing both the goal and safety constraints with a reward function may result in learning a policy that either omits safety or is too conservative; because states that yield high rewards for achieving the goal are often closer to unsafe zones. To overcome this problem, a dual-arm cooperation method is proposed, which decomposes the overall task into two subtasks assigned to each robot arm. This approach ensures that the goal is achieved while considering safety constraints and preventing potential unsafe situations. In this method, while one arm learns to achieve the goal successfully, the other arm learns to interact with the environment to ensure safety by defining an environmental constraint. The proposed method was tested in a simulation environment on 3 different scenarios adapted for a humanoid robot to perform the task of stirring a bowl, and it was seen that the second arm-assisted stirring scenario showed the best performance.