ExBody (Expressive Whole-Body Control for Humanoid Robots)
Robotics modelApache-2.0 Isaac Gym RL whole-body humanoid controller for the 19-DoF Unitree H1 that tracks retargeted CMU Mocap upper-body motions while relaxing lower-body constraints to robust root velocity tracking; upstream ships an in-repo pretrained H1 checkpoint (060-40/model_49500.pt), TorchScript export, and ASE-based motion retargeting scripts. [1] [2] [3] [4]
Setup & usage →CheckpointsModel hub / homepageGetting started
- For
- Humanoid locomotion and whole-body control researchers; Sim-to-real RL practitioners; Unitree H1 developers [1] [2]
- Parameters
- Compact RL MLP (~11.8 MB PyTorch checkpoint for 060-40/model_49500.pt including actor, critic, and optimizer state [1] [2] [5]
- Source rights
- Some licenses are unresolved. Review licenses
Source researched · Not tested by us
Sources checked · Research details
Licenses & related hardware
- License
- Apache-2.0 · license not fully assessed for model weights, datasets · Details by part
Overview
ExBody (*Expressive Whole-Body Control for Humanoid Robots*) is an Apache-2.0 reinforcement learning whole-body controller from UC San Diego designed for the 19-DoF Unitree H1 humanoid. Rather than enforcing strict full-body kinematic imitation that breaks under hardware embodiment gaps, ExBody encourages the upper body to mimic retargeted CMU Mocap reference motions while relaxing lower-body tracking to robust root velocity following. Notably, chengxuxin/expressive-humanoid commits a pretrained Unitree H1 checkpoint (legged_gym/logs/h1/060-40/model_49500.pt, 11.8 MB) directly in the repository alongside training, playback, and TorchScript export scripts. [1] [3] [4] [5]
Model card summary
- Architecture
ActorCriticMimic / PPOMimic RL whole-body controller in RSL-RL/Legged Gym using history-conditioned proprioception (history_len: 10, prop_hist_len: 4), motion demo targets (n_demo_steps: 2), privileged latent features, an optional [128, 64] state estimator, and a distillation student config (H1MimicDistillCfgPPO, [1024, 1024, 512] ELU MLP). Exports traced TorchScript JIT models via save_jit.py.
- Parameter count
Compact RL MLP (~11.8 MB PyTorch checkpoint for 060-40/model_49500.pt including actor, critic, and optimizer state; exact scalar parameter count not stated in README)
Inferred from verified 11,805,165-byte model_49500.pt checkpoint size and H1MimicCfgPPO / H1MimicDistillCfgPPO definitions.
- Input modalities
- proprioceptionimu
Takes 19-DoF joint positions/velocities, IMU/angular velocity proprioception history, and reference upper-body/root velocity demo commands.
- Output modalities
- joint-positions
19-D PD joint position targets (num_policy_actions: 19, control_type: 'P', action_scale: 0.25) for Unitree H1.
- Action space
19-D continuous PD joint position targets (control_type: 'P', action_scale: 0.25, clip_actions: 10, decimation: 4) for Unitree H1 legs, torso, shoulders, and elbows.
- Control frequency
50 Hz (standard Legged Gym sim dt 0.005 s with decimation: 4)
h1_mimic_config.py sets decimation = 4 inheriting LeggedRobotCfg sim dt.
[2] · Editorial assessment
- Training datasets
- CMU Motion Capture (Mocap) FBX dataset retargeted to Unitree H1 (Google Drive link in README)
- Interfaces
- Python CLI scripts (train.py, play.py, save_jit.py)
- Isaac Gym viewer and Flask web viewer (--web)
- TorchScript JIT exported policy checkpoints
Checkpoints
| Checkpoint | Format | Size | Access | Pinned revision |
|---|---|---|---|---|
| ExBody Unitree H1 pretrained policy (060-40 / model_49500.pt) | pytorch | 0.012 GB | open | 511f49582f43 |
Verified in-repo PyTorch checkpoint (11,805,165 bytes) referenced by README quickstart command `python play.py 060-40 --delay --motion_name motions_debug.yaml`. Repository also contains unrelated upstream ASE character checkpoints in ASE/ase/data/models/.
Is this for you?
- Use cases
- Expressive upper-body motion imitation combined with bipedal walking on Unitree H1
- CMU Mocap FBX retargeting and motion-curriculum RL training in Isaac Gym
- Exporting traced TorchScript JIT policies for real-robot deployment
- Known limitations
- Depends on legacy Python 3.8, PyTorch 1.10.0+cu113, proprietary NVIDIA Isaac Gym binaries, and Autodesk FBX SDK for dataset retargeting.
- Bundles third-party ASE retargeting code (nv-tlabs/ASE, NVIDIA Source Code License) and BSD-3-Clause legged_gym/rsl_rl alongside the Apache-2.0 root license.
- Real-world robot low-level deployment driver is not bundled in this repo; save_jit.py only exports traced TorchScript models.
- Linked Google Drive CMU Mocap FBX archive does not declare a separate dataset license in the repository.
- Safety and limitations
- Policy is trained in simulation on a 19-DoF Unitree H1 URDF; real-robot deployment requires external low-level robot interface code and safety supervision.
- Bundled ASE retargeting subtree carries an upstream NVIDIA Source Code License with separate restrictions from the root Apache-2.0 license.
- Action delay is only introduced after 8k training iterations (--delay required when playing checkpoints past 8k).
Compute & fine-tuning
- Inference compute
- Documented hardware: CUDA GPU or CPU (play.py supports --device cuda:0 or cpu; TorchScript JIT export for onboard humanoid deployment)
- Precision:
fp32 - Edge / consumer capable: Yes
Compact MLP exported via save_jit.py runs on CPU/onboard humanoid compute for real-world deployment.
- Fine-tuning
- Supported: Yes
- Methods: PPOMimic RL training and resume from checkpoint in Isaac Gym (legged_gym/legged_gym/scripts/train.py)
- Recipe: Upstream fine-tuning documentation
- Host computer
Single CUDA GPU for Isaac Gym training (default num_envs=6144, --device cuda:0); playback supports cuda:0 or cpu.
- Operating system
- Linux
Inferred from Isaac Gym Linux binaries and CUDA 11.3 setup.
[1] · Editorial assessment
- Dependencies
- Python 3.8
- PyTorch 1.10.0+cu113
- NVIDIA Isaac Gym
- numpy < 1.24
- Autodesk FBX SDK (for mocap import)
- Bundled ASE, legged_gym, and rsl_rl
- Skills
- Python
- PyTorch / TorchScript
- Isaac Gym / Legged Gym / RSL-RL
- Mocap FBX retargeting
Inferred from README installation, ASE poselib retargeting, and save_jit.py workflow.
[1] · Editorial assessment
Setup & usage
Open the upstream getting-started guide [1] · Model card / documentation [1]
- Create a Python 3.8 conda environment and install PyTorch 1.10.0+cu113, Isaac Gym binaries, and bundled rsl_rl and legged_gym in editable mode
- Install numpy<1.24, dependencies, and Autodesk FBX SDK (for mocap import)
- Download the CMU Mocap FBX archive from Google Drive and run fbx_importer_all.py, retarget_motion_h1_all.py, and h1_view keybody extraction
- Play the bundled pretrained checkpoint (python play.py 060-40 --delay --motion_name motions_debug.yaml), train new runs with train.py, or export TorchScript via save_jit.py
Upstream-reported benchmarks
Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.
| Benchmark | Upstream-reported score | Scope & caveats |
|---|---|---|
| Expressive whole-body motion imitation in Isaac Gym and real-world Unitree H1 (arXiv:2402.16796v2) | Sim-to-real transfer enabling diverse walking styles, handshaking, and dancing with a human on a physical Unitree H1 | Upstream-reported simulation and real-robot evaluation on Unitree H1; not independently verified by OSR. |
Project notes
Decoupled whole-body policy and distillation configs
At the pinned commit (511f49582f43), h1_mimic_config.py configures 6,144 parallel Isaac Gym environments controlling 19 Unitree H1 joints via PD position targets (control_type = 'P', action_scale = 0.25, decimation = 4; the config inherits the base simulation timestep rather than declaring dt locally, so 50 Hz is an inferred policy rate assuming a standard 0.005 s Legged Gym timestep). The ActorCriticMimic / PPOMimic policy consumes proprioceptive history (history_len = 10, prop_hist_len = 4) and motion reference targets (n_demo_steps = 2), alongside an optional [128, 64] privileged state estimator and an H1MimicDistillCfgPPO student distillation config ([1024, 1024, 512] ELU MLP). save_jit.py exports traced TorchScript models for deployment. Exact scalar parameter counts are not reported in the README. [1] [2]
Setup constraints and upstream-reported transfer
The repository pins legacy simulation dependencies: Python 3.8, PyTorch 1.10.0+cu113, proprietary NVIDIA Isaac Gym binaries, numpy<1.24, and the Autodesk FBX SDK for importing the linked Google Drive CMU Mocap archive via ASE's poselib. Upstream provides a quickstart playback command for the bundled checkpoint (python play.py 060-40 --delay --motion_name motions_debug.yaml) and notes that action delay is enabled after 8k training iterations. Low-level physical robot drivers are not included in the repo. In arXiv:2402.16796v2, the authors report sim-to-real transfer on a physical Unitree H1 demonstrating diverse walking styles, handshaking, and dancing with a human. All reported results are upstream-reported; Open Source Robotics has not run training, fine-tuning, inference, or robot evaluations for this profile. [1] [4]
What we checked
None. Prepared from public upstream sources without hands-on training, inference, or robot execution.
Updates
- Upstream last updated
- Mar 30, 2025 [1]
- Latest release
- Not established
- Profile updated
- Nov 21, 2026
- Sources last checked
- Nov 21, 2026
- Next review
- Feb 19, 2027
- Reviewed revision
- chengxuxin/expressive-humanoid@511f49582f432f2081a4eaa070f7e9d5f822fa4c · software repository
Separate license matrix
Apache-2.0 whole-body humanoid control codebase with an in-repo pretrained Unitree H1 checkpoint (model_49500.pt), included as partial because training and retargeting depend on proprietary Isaac Gym, NVIDIA ASE's separate license, Autodesk FBX SDK, and an un-licensed Google Drive CMU Mocap archive.
| Layer | Availability / license |
|---|---|
| Code | available · Apache-2.0 Root LICENSE is Apache-2.0 (Copyright 2024 Xuxin Cheng et al.); bundled ASE subtree retains NVIDIA's ASE license and legged_gym/rsl_rl retain BSD-3-Clause. [1] [3] |
| Model weights | available · License not fully assessed Pretrained Unitree H1 checkpoint (legged_gym/logs/h1/060-40/model_49500.pt, 11.8 MB) is committed inside an Apache-2.0 root repo under the BSD-3-Clause legged_gym subtree without an explicit checkpoint-specific license statement. [1] [3] [5] |
| Training datasets | available · License not fully assessed CMU Mocap FBX zip archive is linked on Google Drive without a separately declared license file in the repository. [1] |
| Documentation | available · Apache-2.0 [1] [3] |
Licensing and artifact boundaries
The root repository and documentation are distributed under the Apache License 2.0 (LICENSE at 511f49582f43, Copyright 2024 Xuxin Cheng et al.). However, bundled ASE/ retargeting code carries NVIDIA's separate upstream ASE license, legged_gym/ and rsl_rl/ carry BSD-3-Clause notices, and while model_49500.pt is committed inside legged_gym/logs/h1/060-40/, upstream does not declare a separate checkpoint-specific license—so applying Apache-2.0 to the weights would be an inference rather than an explicit upstream grant. The linked Google Drive CMU Mocap archive likewise does not declare a separate license file in the repository. [1] [3] [5]
References
Each source links to the exact revision we read. “Used for” shows where this page relies on it.
- S001 chengxuxin/expressive-humanoid GitHub repository and README.md
Documents Python 3.8 / PyTorch 1.10+cu113 / Isaac Gym / FBX setup, CMU Mocap Google Drive link, ASE retargeting, pretrained playback command (060-40), and save_jit.py export.Revision, date & where cited
Xuxin Cheng / UC San Diego (GitHub) · github.com · revision
511f495· README.md at commit 511f49582f43 (2025-03-30) · Accessed Nov 21, 2026
Used for: Overview, Model card summary, Checkpoints, Is this for you? and 6 more - S002 ExBody Unitree H1 environment and PPO configuration (h1_mimic_config.py)
Defines 19-DoF H1 PD control (action_scale: 0.25, decimation: 4, num_envs: 6144), ActorCriticMimic / PPOMimic, [128, 64] estimator, and H1MimicDistillCfgPPO ([1024, 1024, 512] ELU student).Revision, date & where cited
Xuxin Cheng / UC San Diego (GitHub) · github.com · revision
511f495· h1_mimic_config.py at commit 511f49582f43 · Accessed Nov 21, 2026
Used for: Model card summary, Compute & fine-tuning, Decoupled whole-body policy and distillation configs - S003 chengxuxin/expressive-humanoid LICENSE
Apache License 2.0 (Copyright 2024 Xuxin Cheng, Yandong Ji, Junming Chen, Ge Yang, Xiaolong Wang), noting ASE upstream license.Revision, date & where cited
Xuxin Cheng / UC San Diego (GitHub) · github.com · revision
511f495· LICENSE at commit 511f49582f43 · Accessed Nov 21, 2026
Used for: Overview, Is this for you?, Separate license matrix - S004 Expressive Whole-Body Control for Humanoid Robots (arXiv:2402.16796v2)
Presents ExBody decoupled upper-body imitation + lower-body velocity tracking trained via RL in simulation and transferred sim-to-real to a physical humanoid.Revision, date & where cited
arXiv (2402.16796v2) · arxiv.org · arXiv abstract 2402.16796v2 · Accessed Nov 21, 2026
Used for: Overview, Is this for you?, Compute & fine-tuning, Upstream-reported benchmarks and 1 more - S005 Pretrained ExBody H1 checkpoint (legged_gym/logs/h1/060-40/model_49500.pt)
Verified 11,805,165-byte (~11.8 MB) PyTorch checkpoint committed inside the repository for run 060-40.Revision, date & where cited
Xuxin Cheng / UC San Diego (GitHub) · github.com · revision
511f495· GitHub Contents API for model_49500.pt at commit 511f49582f43 · Accessed Nov 21, 2026
Used for: Overview, Model card summary, Checkpoints, Separate license matrix
Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1
Change history
2026-11-21 — Initial 0.8.0-draft.1 model profile drafted from chengxuxin/expressive-humanoid@511f49582f43.