Robust Robot Walker: Learning Agile Locomotion over Challenge Terrains
Shaoting Zhu, Runhan Huang, Linzhan Mou, Hang Zhao
ICRA 2025

Installation

1. Create a Python Virtual Environment

• Create a new Python virtual environment with Python 3.6, 3.7, or 3.8 (Python 3.8 is recommended).

2. Install PyTorch 1.10 with CUDA 11.3

Run the following command to install PyTorch:

pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio==0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

3. Install Isaac Gym

• Download and install Isaac Gym Preview 4 from NVIDIA Isaac Gym.
• Install the required Python bindings:

cd isaacgym/python && pip install -e .

• Test the installation by running an example:

cd examples && python 1080_balls_of_solitude.py

• For troubleshooting, check the Isaac Gym documentation at isaacgym/docs/index.html.

4. Clone This Repository

Clone the repository with the following command:

git clone https://github.com/zst1406217/robust_robot_walker.git
cd robust_robot_walker

5. Install rsl_rl (PPO Implementation)

cd rsl_rl && pip install -e .

6. Install legged_gym

cd legged_gym && pip install -e .

7. Install Additional Dependencies

Install the necessary dependencies:

pip install debugpy tqdm numpy==1.19.5 tensorboard==2.0.0 setuptools==58.5.0 protobuf==3.20.0 matplotlib==3.4.0

8. Download the AMP dataset

Download and put the mpc_data_no_command.npy in ./legged_gym.

Run Tiny-Trap Benchmark

1. Navigate to the legged_gym folder:

cd legged_gym

2. The trained policy is stored in checkpoint.zip. Unzip it and place the folder in legged_gym/logs/rrw_a1.

3. Make sure you're in the robust_robot_walker/legged_gym directory, and run the benchmark:

python legged_gym/scripts/track.py --task a1_bartrack --load_run checkpoint --headless

This will output the success rate, average pass time, and average travel distance of 1,000 robots. Due to the random seed and varying GPU devices, results may slightly differ from the paper.

Policy Training

1. Train Walking Policy on Plane

To train the walking policy on a flat plane:

cd legged_gym

Ensure you are in the robust_robot_walker/legged_gym directory, and then run:

python legged_gym/scripts/train.py --task a1_remotegoal --headless

2. Train Robust-Robot-Walker Policy (Stage 1)

1. Update the "load_run" parameter in the legged_gym/envs/a1/a1_mix_goal_stage1_config.py file (line 182) with the log directory from Step 1 (Train walking policy on the plane).

   For example:

   load_run = "Mar14_12-59-47_WalkByRemoteGoal_noResume"
   

2. Make sure you're in the robust_robot_walker/legged_gym folder and run:

python legged_gym/scripts/train.py --task a1_mixgoalstage1 --headless

3. Train Robust-Robot-Walker Policy (Stage 2)

1. Update the "load_run" parameter in the legged_gym/envs/a1/a1_mix_goal_stage2_config.py file (line 182) with the log directory from Step 2 (Train robust-robot-walker policy, stage 1).

2. Run the following:

python legged_gym/scripts/train.py --task a1_mixgoalstage2 --headless

Test and Visualize Walking Policy on Plane

To test and visualize the walking policy:

cd legged_gym
python legged_gym/scripts/play.py --task a1_mixgoalstage2 --load_run <your_log_dir>

Deployment on Real Robot

Please refer to the Deploy.md for instructions on deploying to a real robot.

Acknowlegement

https://github.com/leggedrobotics/legged_gym
https://github.com/ZiwenZhuang/parkour

Citation

You can find our paper on arXiv.

If you find this code or find the paper useful for your research, please consider citing:

@article{zhu2024robust,
  title={Robust Robot Walker: Learning Agile Locomotion over Tiny Traps},
  author={Shaoting, Zhu and Runhan, Huang and Linzhan, Mou and Hang, Zhao},
  journal={arXiv preprint arXiv:2409.07409},
  year={2024}
}

---