Bach Robot Suite

This repo contains some novel robot simulators with PyBullet with Gym environments for training RL agents.

Usage

First install the dependencies for the suite,

pip install -r --user requirements/requirements.txt

The training examples in the examples/ directory use Hydra and my implementation of SAC. To run these training examples, you can install their dependencies with

pip install -r --user requirements/examples.txt

Then, you may install the robot suite with

python setup.py install --user

This installs two packages, brs_data and brs_envs; the former containing the robot definitions, and the latter containing the Gym environments. If you simply wish to use the environments (as opposed to developing new ones), you only need to concern yourself with the brs_envs package.

Training

You can execute a training run with

python examples/rocket_lander.py

The examples/rocket_lander.py script is configured by examples/config-rocket-lander.yaml using Hydra. You may tune some hyperparameters there or via the command line, for example

python examples/rocket_lander.py training.steps=50000 \
                                 agent.params.max_replay_capacity=100000 \
                                 agent.params.env_kwargs.max_pitch_offset=0.1