Manipulation in various tasks

In recent years, reinforcement learning has been used to control robotic manipulation tasks, especially contact-rich ones. These contact-rich tasks requre precise force control to ensure safe and effective performance. Traditional control methods such as impedance control and admittance control, has proved its effectiveness for many years. However they suffer from poor robustness in complex environments. While reinforcement learning my exhibit insecure preformance in control problems, it has the advantage of direct image utilization and robustness to environments. 

In this paper, a integration of reinforcement learning and conventional control method, admittance impedance control is proposed. The proposed method is validated on a wiping simulation, where continous contact force plays a crucial role in completing the task. Our results indicate that manipulation is executed with more delicacy in terms of how contact force is exerted on objects. As the focus was primarily on minimizing contact force, the efficiency of the operation suffered as a result, leading to a "swooping action". Nevertheless, our method resulted with much less average contact force compared to other studies, demonstrating its effectiveness in contact-rich tasks.

Paper/publication:
This research was conducted as a B.S. thesis paper. You can see my full paper here.

Github repository:
https://github.com/n00Nspr1ng/RL_force_control_manipulation