rl-experiments

Reinforcement learning experiments

status

PPO + Panda sorta works on simple pushing task.

contents

mypanda/: Hack of Panda environment with two diff goals - knocking the object around, or holding close to it

ppo_panda.py: PPO implementation targeting the Panda environment

train.py: framework to iterate (potentially endlessly) over hyperparameters and log results. this is from my "depressed and drifting" period

how to use

Install pre-requisites:

$ pip install -r requirements.txt

Edit train.py to tweak the hyperparameters you want to try out (optional). Specify parameter options as lists, and training will try random combinations.

$ vim train.py

Set the system up to render live by creating /tmp/render.txt. Remove this file to turn off the GUI's rendering, which is faster.

$ touch /tmp/render.txt

Start training! The system will collect statistics in ./train-log.jsonlines.

$ python train.py

Weights will be saved every 1000 episodes to panda_saved.

origins

PyBullet Panda simulation: (unknown)

PPO implementation: https://github.com/wisnunugroho21/reinforcement_learning_ppo_rnd