A legged_gym based framework for training legged robots in genesis
-
Totally based on legged_gym
-
Faster and Smaller
For a go2 walking on the plane task with 4096 envs, the training speed in Genesis is approximately 1.3x compared to Isaac Gym, while the graphics memory usage is roughly 1/2 compared to IsaacGym.
With this smaller memory usage, it's possible to run more parallel environments, which can further improve the training speed.
-
Incorporation of various methods in published RL papers
Method Paper Link Location Materials Periodic Gait Reward Sim-to-Real Learning of All Common Bipedal Gaits via Periodic Reward Composition go2_wtw Comparison between smooth gait function and step gait function System Identification Learning Agile Bipedal Motions on a Quadrupedal Robot go2_sysid Teacher-Student Rapid Locomotion via Reinforcement Learning go2_ts
For tests conducted on Genesis, please refer to tests.md
- Create a new python virtual env with python>=3.10
- Install PyTorch
- Install Genesis following the instructions in the Genesis repo
- Install genesis_lr
git clone [email protected]:lupinjia/genesis_lr.git cd genesis_lr pip install -e .
By default, the task is set to go2(in utils/helpers.py), we can run a training session with the following command:
cd legged_gym/scripts
python train.py --headless # run training without renderingAfter the training is done, paste the run_name under logs/go2 to load_run in go2_config.py:
Then, run play.py to visualize the trained model:
For more detailed instructions, please refer to the wiki page
For deployment code, please refer to go2_deploy
| Robot | Sim | Real |
|---|---|---|
| Unitree Go2 (Walk These Ways) | video | video |
| Bipedal Walker |  |
- Add go2 deploy demos and instructions (2025/07/25)
- Add explicit estimator support
- Add teacher-student implementation (2025/08/02)