This code is a reimplementation of the guided policy search algorithm and LQG-based trajectory optimization, meant to help others understand, reuse, and build upon existing work.
For full documentation, see rll.berkeley.edu/gps.
The code base is a work in progress. See the FAQ for information on planned future additions to the code.
This version adds following features:
- Better TensorFlow support
- Add experiments for training
agent_box2dto reach any goal position from any starting position - Create a UR agent which has stable action publish frequency
- Add support for multithreading sampling for UR agent
- Replace the original GMM with GaussianMixture from sklearn
- Add experiments which can train UR robot to go to any target position
- Add experiments which can train UR robot to go to any target poisition with specified orientation
- Add
AlgorithmSLto support training agent with pure supervised learning without any optimal control, to demonstrate the necessity of optimal control
To see the official code, please visit https://github.com/cbfinn/gps