Human to Humanoid

Official Implementation for H2O and OmniH2O:

• Learning Human-to-Humanoid Real-Time Whole-Body Teleoperation, IROS 2024.
• OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning, CoRL 2024.

This codebase is under CC BY-NC 4.0 license, with inherited license in Legged Gym and RSL RL from ETH Zurich, Nikita Rudin and NVIDIA CORPORATION & AFFILIATES. You may not use the material for commercial purposes, e.g., to make demos to advertise your commercial products.

Please read through the whole README.md before cloning the repo.

Installation

Note: Before running our code, it's highly recommended to first play with RSL's Legged Gym version to get a basic understanding of the Isaac-LeggedGym-RslRL framework.

1. Create environment and install torch

   conda create -n omnih2o python=3.8 
   conda activate omnih2o
   pip3 install torch torchvision torchaudio 
   

2. Install Isaac Gym preview 4 release https://developer.nvidia.com/isaac-gym

   unzip files to a folder, then install with pip:

   cd isaacgym/python && pip install -e .

   check it is correctly installed by playing:

   cd examples && python 1080_balls_of_solitude.py

3. Clone this codebase and install our rsl_rl in the training folder

   pip install -e rsl_rl

4. Install our legged_gym

   pip install -e legged_gym

   Ensure you have installed the following packages:

   • pip install numpy==1.20 (must < 1.24)
   • pip install tensorboard
   • pip install setuptools==59.5.0

5. Install our phc

   pip install -e phc

6. Install additional packages requirements.txt

   pip install -r requirements.txt

Training and Playing

1. Try training and playing privileged teacher policy.

   can use "--headless" to disable gui, press "v" to pause/resume gui play.

   # OmniH2O Training and Playing Teacher Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_TEACHER env.num_observations=913 env.num_privileged_obs=990 motion.teleop_obs_version=v-teleop-extend-max-full motion=motion_full motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5
   
   # OmniH2O Play Teacher Policy
   python  legged_gym/scripts/play_hydra.py --config-name=config_teleop task=h1:teleop env.num_observations=913 env.num_privileged_obs=990 motion.future_tracks=True motion.teleop_obs_version=v-teleop-extend-max-full motion=motion_full  motion.extend_head=True asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=10.0  sim_device=cuda:0 load_run=OmniH2O_TEACHER checkpoint=XXXX num_envs=1 headless=False
   

2. Try training and playing sim2real deploy policy.

   # OmniH2O Distill Student Policy
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT env.num_observations=1665 env.num_privileged_obs=1742 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=True env.short_history_length=25 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   
   # OmniH2O Play Student Policy
   python legged_gym/scripts/play_hydra.py --config-name=config_teleop task=h1:teleop env.num_observations=1665 env.num_privileged_obs=1742 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=1 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=10.0 sim_device=cuda:0 load_run=OmniH2O_STUDENT checkpoint=XXXX env.add_short_history=True env.short_history_length=25 headless=False 
   
   

3. Different Configurations on Hisotry Steps

   0-step MLP

   # OmniH2O Distill 0-step MLP Student Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT_0stepMLP env.num_observations=90 env.num_privileged_obs=167 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=False noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   
   # OmniH2O Play 0-step MLP Student Policy 
   python legged_gym/scripts/play_hydra.py --config-name=config_teleop task=h1:teleop env.num_observations=90 env.num_privileged_obs=167 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=1 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=10.0 sim_device=cuda:0 load_run=OmniH2O_STUDENT checkpoint=XXXX env.add_short_history=False headless=False 
   

   5-step MLP

   # OmniH2O Distill 5-step MLP Student Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT_50stepMLP env.num_observations=405 env.num_privileged_obs=482 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=True env.short_history_length=5 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   
   # OmniH2O Play 5-step MLP Student Policy 
   python legged_gym/scripts/play_hydra.py --config-name=config_teleop task=h1:teleop env.env.env.num_observations=405 env.num_privileged_obs=482 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=1 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=10.0 sim_device=cuda:0 load_run=OmniH2O_STUDENT checkpoint=XXXX env.add_short_history=True env.short_history_length=5 headless=False 
   

   50-step MLP

   # OmniH2O Distill 50-step MLP Student Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT_50stepMLP env.num_observations=3240 env.num_privileged_obs=3317 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=True env.short_history_length=50 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   
   # OmniH2O Play 50-step MLP Student Policy 
   python legged_gym/scripts/play_hydra.py --config-name=config_teleop task=h1:teleop env.env.num_observations=3240 env.num_privileged_obs=3317 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=1 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=10.0 sim_device=cuda:0 load_run=OmniH2O_STUDENT checkpoint=XXXX env.add_short_history=True env.short_history_length=50 headless=False 
   

4. Different Configurations on Hisotry Architectures

   LSTM

   # OmniH2O Distill LSTM Student Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=LSTM_STUDENT env.num_observations=90 env.num_privileged_obs=167 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=False rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=False env.short_history_length=0 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=55500 train.dagger.dagger_only=True train.runner.policy_class_name=ActorCriticRecurrent train.policy.rnn_type=lstm
   

   GRU

   # OmniH2O Distill GRU Student Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=LSTM_STUDENT env.num_observations=90 env.num_privileged_obs=167 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=False rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=False env.short_history_length=0 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=55500 train.dagger.dagger_only=True train.runner.policy_class_name=ActorCriticRecurrent train.policy.rnn_type=gru
   

5. Different Configurations on Observation Type (tracking points)

   # OmniH2O Distill 8-point Tracking Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT_8point env.num_observations=1719 env.num_privileged_obs=1796 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[left_ankle_link,right_ankle_link,left_shoulder_pitch_link,right_shoulder_pitch_link,left_elbow_link,right_elbow_link] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=True env.short_history_length=25 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   
   # OmniH2O Distill 23-point Tracking Policy 
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT_23point env.num_observations=1845 env.num_privileged_obs=1922 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[pelvis,left_hip_yaw_link,left_hip_roll_link,left_hip_pitch_link,left_knee_link,left_ankle_link,right_hip_yaw_link,right_hip_roll_link,right_hip_pitch_link,right_knee_link,right_ankle_link,torso_link,left_shoulder_pitch_link,left_shoulder_roll_link,left_shoulder_yaw_link,left_elbow_link,right_shoulder_pitch_link,right_shoulder_roll_link,right_shoulder_yaw_link,right_elbow_link] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:1 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=True env.short_history_length=25 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   

6. Different Configurations on Observation Type (with linear velocity)

   # OmniH2O Distill Student Policy with Linear Velocity
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_STUDENT env.num_observations=1743 env.num_privileged_obs=1820 motion.teleop_obs_version=v-teleop-extend-vr-max motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 train.distill=True train.policy.init_noise_std=0.001 env.add_short_history=True env.short_history_length=25 noise.add_noise=False noise.noise_level=0 train.dagger.load_run_dagger=TEACHER_RUN_NAME train.dagger.checkpoint_dagger=XXX train.dagger.dagger_only=True
   

7. Different Configurations on Training Pipeline (without DAgger)

   # OmniH2O Train Sim2Real Policy with RL directly
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=OmniH2O_wo_DAgger env.num_observations=1665 env.num_privileged_obs=1742 motion.teleop_obs_version=v-teleop-extend-vr-max-nolinvel motion.teleop_selected_keypoints_names=[] motion.extend_head=True num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 noise.add_noise=False noise.noise_level=0 env.add_short_history=True env.short_history_length=25
   

8. Train H2O Policy (8point tracking, no history, MLP, with linear velocity in the state space)

   # H2O Train Sim2Real Policy (8point tracking, no history, MLP, with linear velocity) with RL directly
   python legged_gym/scripts/train_hydra.py --config-name=config_teleop task=h1:teleop run_name=H2O_Policy env.num_observations=138 env.num_privileged_obs=215 motion.teleop_obs_version=v-teleop-extend-max motion.teleop_selected_keypoints_names=[left_ankle_link,right_ankle_link,left_shoulder_pitch_link,right_shoulder_pitch_link,left_elbow_link,right_elbow_link] motion.extend_head=False num_envs=4096 asset.zero_out_far=False asset.termination_scales.max_ref_motion_distance=1.5 sim_device=cuda:0 motion.motion_file=resources/motions/h1/stable_punch.pkl rewards=rewards_teleop_omnih2o_teacher rewards.penalty_curriculum=True rewards.penalty_scale=0.5 env.add_short_history=False
   

Motion Retargeting

AMASS Dataset Preparation

Download AMASS Dataset with SMPL + H G format and put it under human2humanoid/data/AMASS/AMASS_Complete/:

|-- human2humanoid
   |-- data
      |-- AMASS
         |-- AMASS_Complete 
               |-- ACCAD.tar.bz2
               |-- BMLhandball.tar.bz2
               |-- BMLmovi.tar.bz2
               |-- BMLrub.tar
               |-- CMU.tar.bz2
               |-- ...
               |-- Transitions.tar.bz2

And then cd human2humanoid/data/AMASS/AMASS_Complete extract all the motion files by running:

for file in *.tar.bz2; do
    tar -xvjf "$file"
done

Then you should have:

|-- human2humanoid
   |-- data
      |-- AMASS
         |-- AMASS_Complete 
               |-- ACCAD
               |-- BioMotionLab_NTroje
               |-- BMLhandball
               |-- BMLmovi
               |-- CMU
               |-- ...
               |-- Transitions

SMPL Model Preparation

Download SMPL with pkl format and put it under human2humanoid/data/smpl/, and you should have:

|-- human2humanoid
   |-- data
      |-- smpl
         |-- SMPL_python_v.1.1.0.zip

Then cd human2humanoid/data/smpl and unzip SMPL_python_v.1.1.0.zip, you should have

|-- human2humanoid
   |-- data
      |-- smpl
         |-- SMPL_python_v.1.1.0
            |-- models
               |-- basicmodel_f_lbs_10_207_0_v1.1.0.pkl
               |-- basicmodel_m_lbs_10_207_0_v1.1.0.pkl
               |-- basicmodel_neutral_lbs_10_207_0_v1.1.0.pkl
            |-- smpl_webuser
            |-- ...

Rename these three pkl files and move it under smpl like this:

|-- human2humanoid
   |-- data
      |-- smpl
         |-- SMPL_FEMALE.pkl
         |-- SMPL_MALE.pkl
         |-- SMPL_NEUTRAL.pkl

Retargeting AMASS to specific humanoid robot

We use an 3-step process to retarget the AMASS dataset to specific humanoid embodiments. Taking H1 as an example here

1. Write forward kinematics of H1 in human2humanoid/phc/phc/utils/torch_h1_humanoid_batch.py
2. Fit the SMPL shape that matches the H1 kinematics in human2humanoid/scripts/data_process/grad_fit_h1_shape.py
3. Retarget the AMASS dataset based on the corresponding keypoints between fitted SMLP shape and H1 using human2humanoid/scripts/data_process/grad_fit_h1.py

cd human2humanoid
python scripts/data_process/grad_fit_h1_shape.py

And you should have

|-- human2humanoid
   |-- data
      |-- h1
         |-- shape_optimized_v1.pkl 

Retargetting

cd human2humanoid
python scripts/data_process/grad_fit_h1.py

You should have:

(h2o) tairanhe@tairanhe-PRO-WS-WRX80E-SAGE-SE:~/Workspace/human2humanoid$ python scripts/data_process/grad_fit_h1.py
Importing module 'gym_38' (/home/tairanhe/Workspace/isaacgym/isaacgym/python/isaacgym/_bindings/linux-x86_64/gym_38.so)
Setting GYM_USD_PLUG_INFO_PATH to /home/tairanhe/Workspace/isaacgym/isaacgym/python/isaacgym/_bindings/linux-x86_64/usd/plugInfo.json
2024-07-11 18:35:43,587 - INFO - logger - logger initialized
  0%|                                                                                                                                                                                                                              | 0/15886 [00:00<?, ?it/s]15886 Motions to process
0-AMASS_Complete_MPI_Limits_03101_ulr1b_poses Iter: 0    256.983:   0%|                                                                                                                                                             | 0/15886 [00:01<?, ?it/s

After this retargeting loop done, you should have your embodiment-specific dataset ready.

Downloading Full retargeted motion dataset after feasibility filter:

Download motion file amass_phc_filtered.pkl here, and put it under human2humanoid/legged_gym/resources/motions/h1/amass_phc_filtered.pkl. Make sure your running command overwrites the default motion file by motion.motion_file=resources/motions/h1/amass_phc_filtered.pkl

Real-World Deployment

System Overview

• Robot: Unitree H1 EDU
• Perception: ZED mini Camera
• Onboard Compute: Orin NX (16GB)
• Dexterous Hand: Inspire Hand
• Wrist Motor: DM-J4310-2EC

Deployment Code

• Unitree H1: Unitree H1 SDK
• Inspire Hand: Inspire Hand (Unitree SDK)
• Wrist Motor: Damiao c++ controller
• Vision Pro: Please refer to VisionProTeleop and OpenTelevision
• RGB Pose Estimation: Please refer to HybrIK
• Diffusion Policy: Please refer to Diffusion Policy and HATO
• Odometry: ZED mini SDK

Trouble Shooting

Contact

• Deployment and Policy Learning in Sim: Tairan He, tairanh@andrew.cmu.edu
• Motion Retargeting: Zhengyi Luo, zluo2@cs.cmu.edu

Issues

You can create an issue if you meet any bugs, except:

• If you cannot run the vanilla RSL's Legged Gym, it is expected that you first go to the vanilla Legged Gym repo for help.
• There can be CUDA-related errors when there are too many parallel environments on certain PC+GPU+driver combination: we cannot solve thiss, you can try to reduce num_envs.
• Our codebase is only for our hardware system showcased above. We are happy to make it serve as a reference for the community, but we won't tune it for your own robots.

Citation

This codebase builds upon prior work. Please adhere to the relevant licensing in the respective repositories. If you use this code in your work, please consider citing our works:

@inproceedings{he2024learning,
  title={Learning human-to-humanoid real-time whole-body teleoperation},
  author={He, Tairan and Luo, Zhengyi and Xiao, Wenli and Zhang, Chong and Kitani, Kris and Liu, Changliu and Shi, Guanya},
  journal={arXiv preprint arXiv:2403.04436},
  year={2024}
}

@inproceedings{he2024omnih2o,
  title={OmniH2O: Universal and Dexterous Human-to-Humanoid Whole-Body Teleoperation and Learning},
  author={He, Tairan and Luo, Zhengyi and He, Xialin and Xiao, Wenli and Zhang, Chong and Zhang, Weinan and Kitani, Kris and Liu, Changliu and Shi, Guanya},
  journal={arXiv preprint arXiv:2406.08858},
  year={2024}
}

Also consider citing these prior works that helped contribute to this project:

@inproceedings{luo2023perpetual,
  title={Perpetual humanoid control for real-time simulated avatars},
  author={Luo, Zhengyi and Cao, Jinkun and Kitani, Kris and Xu, Weipeng and others},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={10895--10904},
  year={2023}
}

@inproceedings{rudin2022learning,
  title={Learning to walk in minutes using massively parallel deep reinforcement learning},
  author={Rudin, Nikita and Hoeller, David and Reist, Philipp and Hutter, Marco},
  booktitle={Conference on Robot Learning},
  pages={91--100},
  year={2022},
  organization={PMLR}
}

@inproceedings{cheng2024open,
  title={Open-TeleVision: teleoperation with immersive active visual feedback},
  author={Cheng, Xuxin and Li, Jialong and Yang, Shiqi and Yang, Ge and Wang, Xiaolong},
  journal={arXiv preprint arXiv:2407.01512},
  year={2024}
}

@software{Park_Teleopeation_System_using,
author = {Park, Younghyo},
title = {{Teleopeation System using Apple Vision Pro}},
url = {https://github.com/Improbable-AI/VisionProTeleop},
version = {0.1.0}
}

@article{peng2018deepmimic,
  title={Deepmimic: Example-guided deep reinforcement learning of physics-based character skills},
  author={Peng, Xue Bin and Abbeel, Pieter and Levine, Sergey and Van de Panne, Michiel},
  journal={ACM Transactions On Graphics (TOG)},
  volume={37},
  number={4},
  pages={1--14},
  year={2018},
  publisher={ACM New York, NY, USA}
}


@article{lin2024learning,
   author={Lin, Toru and Zhang, Yu and Li, Qiyang and Qi, Haozhi and Yi, Brent and Levine, Sergey and Malik, Jitendra},
   title={Learning Visuotactile Skills with Two Multifingered Hands},
   journal={arXiv:2404.16823},
   year={2024}
}