Comparison of Control Methods: Learning Robotics Manipulation with Contact Dynamics

We compare the different control methods in learning a robotic manipulation task. The task is to push an object (a cube and sphere) from varying beginning position to a fixed goal position. Complex contact dynamics is involved. We used PPO as the learning algorithm trained from scratch with dense rewards. Comparison is performed on two dimensions: learning at joint level vs. end-effector level, as well as velocity control vs. position control. For end-effector learning, we use inverse jacobian to map from end-effector target velocity/position to joint velocity/position, and accounting for singularity, joint limits, and gimbal lock. Across the four methods proposed, joint velocity control demonstrated the fastest convergence on cube task across all control methods, and is the only successful method on sphere task. Video demonstration: https://www.youtube.com/watch?v=wh qV58f95Y