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

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 git@github.com: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

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).

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.

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

Play distillation policy:

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

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.

Tips

3 pre-trained models are provided in legged_gym/logs/parkour_new. You can play with them directly.

051-40: base policy with scandots as input.
051-41: distillation policy with depth as input. No heading direction distillation.
051-42: distillation policy with depth as input. With heading direction distillation.

If you don't need direction distillation, comment out the following line in rsl_rl/rsl_rl/runners/on_policy_runner.py:

obs_student[infos["delta_yaw_ok"], 6:8] = yaw.detach()[infos["delta_yaw_ok"]]

and line in legged_gym/legged_gym/scripts/play.py:

obs[:, 6:8] = 1.5*yaw

Deployment

A1 + internal realsense D435i + Jetson Xavier NX.
Hardware code and Go1 support coming later.

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}
}

dkanou/extreme-parkour