This repository is a framework used to train RL agents in a Panda environment [1]. The main purpose of it is to test different algorithms on different tasks to solve complex problems in a 3-dimensional space.
git clone https://github.com/trjohnny/RL_project.git
py main.py --env [environment] --algo [algorithm] --verbose [1,2 or 3] --episodes [episodes]
- environment: listed here ----> https://github.com/qgallouedec/panda-gym
- algo: A2C (Actor Critic) | A2C_DISCRETE (Actor Critic with discrete actions aggregations) | DDPG (Deep Deterministic Policy Gradient) | A2C_N_STEP_AHEAD (Actor Critic TD(n) )
- verbose: 1 (print every 100 episodes) | 2 (every 10) | 3 (print every episode)
- episodes: the number of episodes to run
Please note: To test the repo with the notebook and change hyperparameters, pass them to the constructor of the Agent.
Hyperparameter optimization is still under review but will be ready soon :)
[1] Gallouédec, Quentin and Cazin, Nicolas and Dellandréa, Emmanuel and Chen, Liming
panda-gym: Open-Source Goal-Conditioned Environments for Robotic Learning
4th Robot Learning Workshop: Self-Supervised and Lifelong Learning at NeurIPS, 2021