Latest News

• [2026.10.24] 🚀 We now support SuperMimic in HoloMotion! Train G1 motion trackers from scratch on a single GPU, with MuJoCo Warp or Isaac PhysX. We achieved 80% SR on LAFAN1 in 30 minutes on a single RTX 5090.

  

• [2026.10.22] The H-bot Competition is open! HoloBoxing learns boxing with about 7 minutes of additional single-GPU training above frozen HoloMotion; see the Track 2 tutorial.

• [2026.10.12] HoloMotion v2.0 brings one policy for sparse-to-whole-body control, with initial 1-point and 5-point real-robot teleoperation demonstrated.

Why HoloMotion?

Scales Toward 4-Any Humanoid Control

HoloMotion progresses from general motion tracking in v1 to flexible command following in v2, then toward cross-embodiment generalization in v3 and terrain generalization in v4.

Version	Target Capability	Status	Focus
v1	Any Pose	✅ Released	Scale motion tracking with large pretrained models, fast inference, and generalization to unseen motions and real robots.
v2	Any Command	🚀 Preview	Unify sparse-to-whole-body commands in one policy, with an SMPL interface for cross-modal extensions.
v3	Any Embodiment	🧭 Planned	Generalize motion and command following across humanoids with different morphologies and actuation.
v4	Any Terrain	🧭 Planned	Adapt whole-body control to uneven and dynamic terrain, maintaining balance and coordinated motion in complex environments.

HoloMotion v1: Larger, Faster and Stronger as a Motion Tracker

HoloMotion v1 uses a reference-conditioned MoE Transformer to track diverse human motions on humanoid robots.

• Larger: Up to 0.4B parameters trained on 2,000+ hours of motion, combining video-derived data with motion capture.
• Faster: Sequence-level training and KV-cache inference support efficient learning and real-time control.
• Stronger: Generalization to unseen motion datasets and real-robot transfer without task-specific fine-tuning, as documented in the HoloMotion-1 Technical Report.

HoloMotion v2: Flexible Sparse-to-Whole-Body Command Following with Cross-Modal Extensibility

• One policy, multiple command modes: Pelvis-1 and Keypoint-5 (pelvis, wrists, and ankles) use the same model with selectable control-point masks.
• Cross-modal extensibility through SMPL: The shared SMPL interface is designed to simplify integration of external text-to-motion and music-to-motion models; these adapters are not yet included.
• Live teleoperation: Initial 1-point and 5-point PICO VR trials have been completed on G1. See deployment.

v2.0 is in release preparation; real-robot gait and tracking quality are still being refined, especially for pelvis-only control. Validation coverage varies across modes.

One End-to-End Framework: From Motion Data to Robot Control

Motion Conversion → Retargeting → Policy Training → Simulation Evaluation → Robot Deployment

Unifies Diverse Motion Sources

HoloSMPL standardizes VR, inertial, optical, and vision capture into a shared representation. HoloRetarget converts these motions into robot references for data generation and live teleoperation.

No Per-User Training Required

Start with the v2.0.0 motion-tracking checkpoint for simulation and live control; train or fine-tune only when you need to adapt the policy.

Goal	Guide	Starting Point
Replay a motion in MuJoCo	Sim2sim	ONNX/config bundle and a robot NPZ clip.
Evaluate a motion dataset	Quantitative evaluation	ONNX/config bundle, actor weights, and robot HDF5 data.
Validate and test a boxing policy	HoloBoxing local evaluation	Your high-level boxing policy (ONNX), the bundled evaluation image, and a compatible GPU.
Teleoperate a G1	Robot deployment	G1 29-DOF with Orin, matching deployment image, and PICO / XRoboToolkit.
Train or fine-tune	Training	Labeled robot HDF5 data and GPU resources; a compatible checkpoint for fine-tuning.
Train a lightweight motion tracker from scratch	SuperMimic	G1 reference motions, one CUDA GPU, and the selected simulator environment.

For workstation workflows, begin with environment setup. Each guide specifies the required artifacts and commands.

Projects Using HoloMotion

Organization	Project	Use of HoloMotion	Links
Tsinghua University · MARS Lab	OMG	Uses HoloMotion as the tracker for executing generated motions, with G1 deployment workflows.	Paper / Project / Code
Peking University · School of Computer Science	HumanoidTTT	Keeps HoloMotion frozen as the whole-body tracker for reusing validated motions.	Paper / Project / Code
RoboParty Lab	sim2real	Supports the official HoloMotion v1.4.0 ONNX policy in its simulation and robot deployment stack.	Code / Docs
ccrpRepo	RoboMimic Deploy	Implements HoloMotion as a selectable G1 policy, with ONNX inference and motion or velocity tracking modes.	Code / HoloMotion policy
Renforce Dynamics	GenesisLab / genPiHub	Integrates HoloMotion v1.2 into Genesis simulation. Velocity tracking is implemented; motion tracking is still in progress.	GenesisLab / genPiHub

Citation

The HoloMotion-2 Technical Report is scheduled for release alongside HoloMotion v2.1 at the end of October 2026. In the meantime, please cite the HoloMotion-1 report:

@misc{chen2026holomotion1,
  title = {HoloMotion-1 Technical Report},
  author = {Maiyue Chen and Kaihui Wang and Bo Zhang and Xihan Ma and Zhiyuan Yang and Yi Ren and Qijun Huang and Zihao Zhu and Yucheng Wang and Zhizhong Su},
  year = {2026},
  eprint = {2605.15336},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2605.15336}
}

For the HoloBoxing method and boxing experiments, please cite the HoloBoxing paper:

@misc{ma2026holoboxing,
  title = {{HoloBoxing}: A Simple Recipe for Humanoid Boxing with 7 Minutes of Single-GPU Training},
  author = {Xihan Ma and Bo Zhang and Boxiong Li and Yuhan Bao and Ziyue Liu and Yucheng Wang},
  year = {2026},
  eprint = {2610.32909},
  archivePrefix = {arXiv},
  primaryClass = {cs.RO},
  url = {https://arxiv.org/abs/2610.32909}
}

Acknowledgements

Built on and inspired by these open-source projects:

• ASAP
• BeyondMimic
• GMR
• GVHMR
• Humanoidverse
• Mink
• MotionMillion
• PBHC
• PHC
• ProtoMotion
• RoboStriker
• ScaleBFM
• SONIC
• SuperMimic [1] [2].