Learning to Cooperate: A Hierarchical Cooperative Dual Robot Arm Approach for Underactuated Pick-and-Placing

Cooperative multiagent manipulation systems allow to extend on the manipulative limitations of individual agents, increasing the complexity of the manipulation tasks the ensemble can handle. Controlling such a system requires meticulous planning of subsequent subtasks, queried to the individual agents, in order to execute the master task successfully. Real-time (re)planning is essential to ensure the task can still be achieved when subtasks execution suffers from uncertainty or when the master task changes intermittently requiring real-time reconfiguration of the plan. In this article, we develop a supervisory control architecture tailored to the cooperation of two robotic manipulators equipped with standard pick-and-place facilities in the plane. We control the planar position and orientation of an object using two underactuated manipulators so that only the position of the object can be controlled directly. The desired orientation follows from the accumulation of alternating relative angles. A time-invariant policy function is trained using deep reinforcement learning, which can determine a finite sequence of pick-and-place maneuvers to manipulate the object to a desired configuration. Two policy architectures are compared. The first uses the kinematic model to determine the final step, whilst the second policy makes this decision itself. The more information is given to the policy, the easier it trains. In return, it becomes less adaptable and loses some of its generalizability.