Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots

[Website] [Paper] [Video]

TODO

• Release motion capture data
• Release BBC Training Pipeline
• Release TSC-teacher Training Pipeline
• Release TSC-student Training Pipeline
• Release EASI Training Pipeline
• Release motion retargeting pipeline
• Release deployment pipeline

Installation

• Create a new python virtual environment with python 3.8

conda create -n agility python=3.8
conda activate agility

• Install pytorch 2.3.1 with cuda-11.8

pip install torch==2.3.1+cu118 torchvision==0.18.1+cu118 torchaudio==2.3.1+cu118 -f https://download.pytorch.org/whl/cu118/torch_stable.html

• Install Isaac Gym
  • Download and install Isaac Gym Preview 4

    cd isaacgym/python
    pip install -e .

  • Try running an example

    cd examples
    python 1080_balls_of_solitude.py

  • For troubleshooting, check docs in isaacgym/docs/install.html
• Clone this repository and install the additional dependencies

git clone https://github.com/NJU-RLC/quadrupedal-agility.git
cd quadrupedal-agility
pip install -r requirements.txt

Usage

Train BBC

Go to bbc/legged_gym/scripts and run

python train.py --task go2_locomotion --experiment_idx 0 --device_id 0 --headless True

• Train ~200k iterations (~4 days on 3090). The trained policy is saved in bbc/logs/go2_locomotion/0/model.pt.
• Use --resume --load_run 0 to start training from a saved checkpoint.

Train TSC

Before training, please copy the trained BBC model model.pt to tsc/weights/bbc (A pre-trained model has been provided). Then go to tsc/legged_gym/scripts and run

# Train TSC-teacher
python train.py --proj_name agility --exptid base --device_id 0 --randomize_start --randomize_base_mass --randomize_base_com --push_robots
# Train TSC-student
python train.py --proj_name agility --exptid dist --device_id 0 --resume --resumeid base --use_camera

• Train ~30k iterations (~20 hours on 3090) for TSC-teacher, and ~15k iterations (~17 hours on 3090) for TSC-student. The trained policies are saved in tsc/logs/agility/base/model.pt and tsc/logs/agility/dist/model.pt.
• Use --resume --resumeid base/dist to start training from a saved checkpoint.

Train EASI

The EASI-optimized simulator parameter settings are adopted for BBC training. More details about EASI can be found at EASI.

Play BBC

Go to bbc/legged_gym/scripts and run

python play.py --task go2_locomotion --load_run 0

• Continuously press W/S for acceleration and deceleration, control steering with A/D, and switch behavior modes between 1-5.

Play TSC

Go to tsc/legged_gym/scripts and run

# Play TSC-teacher
python play.py --proj_name agility --exptid base
# Play TSC-student
python play.py --proj_name agility --exptid dist --use_camera

• Switch to next/prev robot with [ ], switch between free camera and following camera with F.

Citation

If you found any part of this code useful, please consider citing:

@article{fu2025learning,
      title={Learning Diverse Natural Behaviors for Enhancing the Agility of Quadrupedal Robots}, 
      author={Huiqiao Fu and Haoyu Dong and Wentao Xu and Zhehao Zhou and Guizhou Deng and Kaiqiang Tang and Daoyi Dong and Chunlin Chen},
      year={2025},
      eprint={2505.09979},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2505.09979}
}

References

The code is built upon the open-sourced IsaacGymEnvs, rsl_rl, legged_gym, MetalHead, and extreme-parkour.