PhysBrain 1.0 is a physical intelligence system built around the paradigm shift from action imitation to physical commonsense acquisition in embodied AI. At its core is the world's first scalable physical intelligence data engine capable of converting large-scale human egocentric video into high-quality multimodal training data — eliminating the dependency on expensive closed-loop robot demonstrations.
The data engine processes over 3,000 hours of human video with fine-grained annotations across spatial relationships, action feasibility, and multi-step logical reasoning in real 3D environments. This corpus moves beyond simple action replication to extract the underlying physical laws and commonsense logic embedded in everyday human activity. When injected into a multimodal large model, it successfully elicits human-like physical intelligence, enabling the model to understand physics rather than merely imitate motions.
Built upon this foundation, this repository provides the Vision-Language-Action (VLA) model for robot control — the bridge from physical intelligence to real-world robotic applications.
Key Technologies
PhysBrain Data Engine
A scalable pipeline that transforms raw human egocentric video into structured, multimodal embodied training data — annotated with spatial structure, motion feasibility, and causal reasoning chains.
TwinBrainVLA — Dual-Brain Fusion Architecture
A novel architecture that addresses the industry-wide challenge of catastrophic forgetting during embodied fine-tuning. By maintaining a parallel general-purpose brain alongside a task-specific embodied brain, TwinBrainVLA achieves "generalist-specialist fusion".
LangForce — Physics-Grounded Training Strategy
A principled training methodology that breaks the visual shortcut dilemma in VLA learning through a Bayesian statistical lens. LangForce fundamentally shifts the training objective from behavioral cloning to physical commonsense acquisition.
Open Source Plan
All PhysBrain 1.0 VLA model checkpoints are now available. You can find the full collection at 🤗 Hugging Face.
Load and deploy the model following the standard starVLA checkpoint loading workflow.
For detailed starVLA setup and inference instructions, please refer to the starVLA repository.
Citation
If you find PhysBrain useful in your research, please consider citing our work:
bibtex
1@misc{physbrain,
2 title={PhysBrain: Human Egocentric Data as a Bridge from Vision Language Models to Physical Intelligence},
3 author={Xiaopeng Lin and Shijie Lian and Bin Yu and Ruoqi Yang and Zhaolong Shen and Changti Wu and Yuzhuo Miao and Yurun Jin and Yukun Shi and Jiyan He and Cong Huang and Bojun Cheng and Kai Chen},
4 year={2026},
5 eprint={2512.16793},
6 archivePrefix={arXiv},
7 primaryClass={cs.RO},
8 url={https://arxiv.org/abs/2512.16793},
9}
1011@misc{physbrain10technicalreport,
12 title={PhysBrain 1.0 Technical Report},
13 author={Shijie Lian and Bin Yu and Xiaopeng Lin and Changti Wu and Hang Yuan and Xiaolin Hu and Zhaolong Shen and Yuzhuo Miao and Haishan Liu and Yuxuan Tian and Yukun Shi and Cong Huang and Kai Chen},
14 year={2026},
15 eprint={2605.15298},
16 archivePrefix={arXiv},
17 primaryClass={cs.RO},
18 url={https://arxiv.org/abs/2605.15298},
19}
