- [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.
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.
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.x | Any Pose | ✅ Done | Achieve robust tracking and imitation of diverse, whole-body human motions, forming the core of the imitation learning capability. |
| v2.x | Any Command | 🔜 Next | Enable language- and task-conditioned motion generation, allowing for goal-directed and interactive behaviors. |
| v3.x | Any Terrain | ⏳ Not started | Master adaptation to uneven, dynamic, and complex terrains, enhancing real-world operational robustness. |
| v4.x | Any Embodiment | ⏳ Not started | Generalize control policies across humanoids with varying morphologies and kinematics, achieving true embodiment-level abstraction. |
HoloMotion provides a clear, modular framework for bridging motion data, policy learning, simulation evaluation, and real-robot deployment.
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 | A pretrained policy and retargeted .npz motion clips. No model training is required. |
| Online motion tracking Follow live VR or teleoperation motion streams. |
Real-world deployment: Online Motion | A pretrained policy, 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 → Data curation → Retargeting → Training → Evaluation | Training environment, curated motion data, retargeted HDF5 datasets, and GPU resources. |
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.
@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}
}
This project is built upon and inspired by several outstanding open source projects: