RL + PyElastica for a Tendon-Driven Continuum Robot

[gabotuzl]

linkedin_RL_cr_robot.mp4

Hello,

Wanted to share my project where I used Gymnasium + SB3 + PyElastica in conjunction with a tendon forcing module I created myself to carry out the control of a tendon driven robot using reinforcement learning (PPO).

At first I tried to get an optimal policy by training the agent from scratch, but it proved to be far more difficult to achieve as local optima are everywhere. So I opted for using a behavioral cloning method where I pretrained the agent's neural network to match the outputs of a simple PID controller which by itself does a decent job at control.
Upon training, the agent was able to surpass the PID controller's metrics and thus achieve better policies.

Let me know if you have any comments or questions! The project is in this repo.