Joint Policy Optimization for Task-Oriented Grasping with a Humanoid Gripper

Reinforcement learning (RL) algorithms signifi-cantly impact robotics by overcoming the limitations of conventional robot control and enabling new-generation collaborative and bio-inspired robots to explore and learn from experience. As a result, we can approach fundamental challenges in robotics such as object grasping and manipulation in a new way and embed their coherent task-oriented relation as an optimization problem. We have trained two RL policies in a self-supervised manner, where the first agent learns to optimally grasp an object based on the success rate of the second agent, which executes the following manipulation task with the previously grasped object. The agents learned to control and manipulate a robotic arm with a humanoid gripper to grasp a hammer and hit a nail in a simulation environment. As a result, the trained agents successfully grasp the hammer in a task-oriented manner and hit the nail in more than 99 % of all given situations. In addition, splitting the tasks resulted in significant training steps reduction even for more than 50 %.