Booster Gym

Booster Gym is a reinforcement learning (RL) framework designed for humanoid robot locomotion developed by Booster Robotics.

News : We now provide a new RL pipeline supporting K1 robot! - Booster Train trains a set of reinforcement learning tasks for Booster robots using Isaac Lab. - Booster Deploy provides an easy-to-use deployment framework that enables seamlessly running the same policy code in both simulation and on real robots. - Booster Assets provides Booster robot descriptions and example motion data.

- Complete Training-to-Deployment Pipeline: Full support for training, evaluating, and deploying policies in simulation and on real robots. - Sim-to-Real Transfer: Including effective settings and techniques to minimize the sim-to-real gap and improve policy generalization. - Customizable Environments and Algorithms: Easily modify environments and RL algorithms to suit a wide range of tasks. - Out-of-the-Box Booster T1 Support: Pre-configured for quick setup and deployment on the Booster T1 robot.

The framework supports the following stages for reinforcement learning:

- Train reinforcement learning policies using Isaac Gym with parallelized environments.

- In-Simulation Testing: Evaluate the trained policy in the same environment with training to ensure it behaves as expected. - Cross-Simulation Testing: Test the policy in MuJoCo to verify its generalization across different environments.

- Model Export: Export the trained policy from .pth to a JIT-optimized .pt format for efficiency deployment - Webots Deployment: Use the SDK to deploy the model in Webots for final verification in simulation. - Physical Robot Deployment: Deploy the model to the physical robot using the same Webots deployment script.

sh $ conda create --name <envname> python=3.8 $ conda activate <envname>

sh $ conda install numpy=1.21.6 pytorch=2.0 pytorch-cuda=11.8 -c pytorch -c nvidia

sh $ tar -xzvf IsaacGymPreview4Package.tar.gz $ cd isaacgym/python $ pip install -e .