bicycle_rl_control

This repository contains a reinforcement learning-based controller for bicycle balancing and navigation in simulation. The project uses the Genesis simulator and implements PPO (Proximal Policy Optimization) for training bicycle control policies.

genesis_bicycle.mp4

Installation

At first, install Genesis via PyPI:

pip install genesis-world

This project use the PPO implementation from rsl-rl to train the policy. Follow these installation steps:

# Install rsl_rl.
git clone https://github.com/leggedrobotics/rsl_rl
cd rsl_rl && git checkout v1.0.2 && pip install -e .

# Install tensorboard.
pip install tensorboard

Training

Train the bicycle policy using the BicycleEnv environment.

Run with:

python scripts/bicycle_train.py -e bicycle-policy -B 8192 --max_iterations 300

Train with visualization:

python scripts/bicycle_train.py -e bicycle-policy -B 8192 --max_iterations 300 -v

Evaluation

Evaluate the trained bicycle policy.

Run with:

python scripts/bicycle_eval.py -e bicycle-policy --ckpt 300 --record

Note: If you experience slow performance or encounter other issues during evaluation, try removing the --record option.

Reference

• 3d model Autonomous bicycle (Apache License 2.0)