du 对 gitlink 路径按磁盘目录递归计了 117M,实际索引里只有 40 字节子模块 指针(mode 160000),STL 从未入库。恢复指针保持 upstream subtree 树形 与官方一致(subtree pull 依赖)。 Co-Authored-By: Claude Code <noreply@anthropic.com>
News
- [2026.07.16] HoloMotion v1.4 introduces HoloRetarget at 3,000+ FPS on RTX 4090 for training data generation and 300+ FPS entirely on-robot for teleoperation, plus HoloSMPL for 10+ datasets and devices.
- [2026.05.15] HoloMotion v1.3 scales from 60M to 0.4B parameters and 80 to 2000+ hours of motion data, while improving policy inference from ~100 to ~300 FPS.
- [2026.04.04] HoloMotion v1.2 provides pre-trained motion tracking and velocity tracking models for the community to deploy directly.
Why HoloMotion?
Larger, Faster and Stronger
HoloMotion scales humanoid whole-body control through a reference-conditioned MoE Transformer, large-scale motion data, and an optimized training-to-deployment pipeline, delivering stronger motion tracking with real-time inference efficiency.
Scales Toward 4-Any Humanoid Control
The roadmap of HoloMotion advances through four generalization targets, from motion imitation to command following, terrain adaptation, and embodiment transfer.
| Version | Target Capability | Status | Description |
|---|---|---|---|
| v1 | Any Pose | ✅ Done | Achieve robust tracking and imitation of diverse, whole-body human motions, forming the core of the imitation learning capability. |
| v2 | Any Command | 🚀 Next | Enable language- and task-conditioned motion generation, allowing for goal-directed and interactive behaviors. |
| v3 | Any Embodiment | 🧭 Planned | Generalize control policies across humanoids with varying morphologies and kinematics, achieving true embodiment-level abstraction. |
| v4 | Any Terrain | 🧭 Planned | Master adaptation to uneven, dynamic, and complex terrains, enhancing real-world operational robustness. |
Unifies Diverse Motion Sources
HoloSMPL standardizes VR, inertial, optical, and vision capture into a shared representation for HoloRetarget.
One Policy Across Modular G1 Configurations
HoloMotion runs one shared motion policy across 2 heads, 5 G1 29-DoF body modes, and 7 hands, covering 66 unique compatible configurations without training a separate policy for every hardware assembly.
The five body modes retain the same 29-DoF kinematic layout while varying pelvis and wrist geometry together with hip, wrist, and waist actuator specifications. The supported modules span Normal and COMP heads; Mode 5, 11, 13, 15, and 18 bodies; and None, Rubber, Dex1-1, Dex3-1, Inspire DFQ, Inspire FTP, and BrainCo Revo2 hand options.
Closes the Loop From Data to Robots
HoloMotion provides a clear, modular framework for bridging motion data, policy learning, simulation evaluation, and real-robot deployment.
No Per-User Training Required
Whether you want to replay motions, stream live teleoperation, or train a custom policy, HoloMotion provides a direct path into the workflow:
| User Goal | Start Here | What You Need |
|---|---|---|
| Offline motion tracking Replay local motion clips for demos such as dance or scripted performances. |
Real-world deployment: Offline Motion | The v1.4.1 Docker image and retargeted .npz motion clips. No model training is required. |
| Online motion tracking Follow live VR or teleoperation motion streams. |
Real-world deployment: Teleoperation | The v1.4.1 Docker image, robot deployment setup, and a live motion source. No model training is required. |
| Train your own model Build a custom policy from your own motion data. |
Environment setup → HoloSMPL conversion → Retargeting → Training → Evaluation | Training environment, curated motion data, retargeted HDF5 datasets, and GPU resources. |
Join Us
We are hiring full-time engineers, new graduates, and interns who are excited about humanoid robots, motion control, and embodied intelligence. Send your resume by scanning the WeChat QR code below to get in touch with us.
Projects Using HoloMotion
We are glad to see HoloMotion being used as a motion control foundation for humanoid research and applications.
| Project | Description | Links |
|---|---|---|
| OMG: Omni-Modal Motion Generation for Generalist Humanoid Control | Uses HoloMotion as the motion control foundation for omni-modal motion generation and generalist humanoid control. | Paper / Project / Code |
| HoloAgent-0: A Unified Embodied Agent Framework with 3D Spatial Memory | Uses HoloMotion as the humanoid whole-body motion skill within its closed-loop Embodied AgentOS. | Paper / Project / Code |
Citation
@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}
}
Acknowledgements
This project is built upon and inspired by several outstanding open source projects:






