Reinforcement learning for controls

Source: https://www.youtube.com/watch?v=aTDkYFZFWug

Jemin Hwangbo et al. ,Learning agile and dynamic motor skills for legged robots.Sci. Robot.4,eaau5872(2019).DOI:10.1126/scirobotics.aau5872

The above video and paper shows what seems to be a basically canonical example of using machine learning for robot controls.

Intro (2021 recitation): https://fab.cba.mit.edu/classes/865.21/topics/control/07_reinforcement_learning.html

Policy Search: https://underactuated.csail.mit.edu/rl_policy_search.html

REINFORCE and more: https://manipulation.csail.mit.edu/rl.html (specifically, this chapter includes an interactive notebook with REINFORCE implemented for a box flip-up task).

Gym training: https://gymnasium.farama.org/

RL algorithms: https://stable-baselines3.readthedocs.io/en/master/index.html (https://jmlr.org/papers/volume22/20-1364/20-1364.pdf)

Learning Reinforcement Learning resources as compiled by stable-baselines3: https://stable-baselines3.readthedocs.io/en/master/guide/rl.html

Is RL a solve all ? no 😛.

Please see [insert repo link] for a drake-gym implementation of the pendubot (as of 04/17/24 model has learned only to either wiggle or crash the simulation, will continue to update).

Erik also has put together great resources for the cartpole control task using a policy gradient method. As well as a PINNs version: Physics informed neural networks.