Extreme Parkour with Legged Robots

Authors: Xuxin Cheng*, Kexin Shi*, Ananye Agarwal, Deepak Pathak
Website: https://extreme-parkour.github.io
Paper: https://arxiv.org/abs/2309.14341
Tweet Summary: https://twitter.com/pathak2206/status/1706696237703901439

Installation

conda create -n parkour python=3.8
conda activate parkour
cd
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
git clone [email protected]:chengxuxin/extreme-parkour.git
cd extreme-parkour
# Download the Isaac Gym binaries from https://developer.nvidia.com/isaac-gym 
# Originally trained with Preview3, but haven't seen bugs using Preview4.
cd isaacgym/python && pip install -e .
cd ~/extreme-parkour/rsl_rl && pip install -e .
cd ~/extreme-parkour/legged_gym && pip install -e .
pip install "numpy<1.24" pydelatin wandb tqdm opencv-python ipdb pyfqmr flask

Usage

cd legged_gym/scripts

1. Train base policy:

python train.py --exptid xxx-xx-WHATEVER --device cuda:0

Train 10-15k iterations (8-10 hours on 3090) (at least 15k recommended).

2. Train distillation policy:

python train.py --exptid yyy-yy-WHATEVER --device cuda:0 --resume --resumeid xxx-xx --delay --use_camera

Train 5-10k iterations (5-10 hours on 3090) (at least 5k recommended).

You can run either base or distillation policy at arbitary gpu # as long as you set --device cuda:#, no need to set CUDA_VISIBLE_DEVICES.

3. Play base policy:

python play.py --exptid xxx-xx

No need to write the full exptid. The parser will auto match runs with first 6 strings (xxx-xx). So better make sure you don't reuse xxx-xx. Delay is added after 8k iters. If you want to play after 8k, add --delay

4. Play distillation policy:

python play.py --exptid yyy-yy --delay --use_camera

5. Save models for deployment:

python save_jit.py --exptid xxx-xx

This will save the models in legged_gym/logs/parkour_new/xxx-xx/traced/.

Viewer Usage

Can be used in both IsaacGym and web viewer.

• ALT + Mouse Left + Drag Mouse: move view.
• [ ]: switch to next/prev robot.
• Space: pause/unpause.
• F: switch between free camera and following camera.

Arguments

• --exptid: string, can be xxx-xx-WHATEVER, xxx-xx is typically numbers only. WHATEVER is the description of the run.
• --device: can be cuda:0, cpu, etc.
• --delay: whether add delay or not.
• --checkpoint: the specific checkpoint you want to load. If not specified load the latest one.
• --resume: resume from another checkpoint, used together with --resumeid.
• --seed: random seed.
• --no_wandb: no wandb logging.
• --use_camera: use camera or scandots.
• --web: used for playing on headless machines. It will forward a port with vscode and you can visualize seemlessly in vscode with your idle gpu or cpu. Live Preview vscode extension required, otherwise you can view it in any browser.

Acknowledgement

https://github.com/leggedrobotics/legged_gym
https://github.com/Toni-SM/skrl

Citation

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

@article{cheng2023parkour,
title={Extreme Parkour with Legged Robots},
author={Cheng, Xuxin and Shi, Kexin and Agarwal, Ananye and Pathak, Deepak},
journal={arXiv preprint arXiv:2309.14341},
year={2023}
}