The Year in AI Papers
The year/Independent research

Paper 2605.15298

PhysBrain 1.0 Technical Report

Shijie Lian, Bin Yu, Xiaopeng Lin, Changti Wu, Hang Yuan, Xiaolin Hu, Zhaolong Shen, Yuzhuo Miao, Haishan Liu, Yuxuan Tian, Yukun Shi, Cong Huang, et al.

Published
May 2026
Research lab
Independent
Citations
0
GitHub
37 stars

01 In brief

Summary

PhysBrain 1.0 introduces a training strategy for embodied AI that prioritizes physical commonsense acquisition over pure action imitation.

Its data engine converts large-scale human egocentric video (Ego4D, BuildAI, EgoDex, EPIC, SEA-Small) into structured scene meta-information—scene elements, spatial dynamics, action execution, and depth-aware relations—which is then rendered into physically grounded question-answer supervision for training a base VLM.

This physically informed base model is adapted to robot control via a dual-pathway architecture that preserves general multimodal capability, an action-conditioned language alignment objective to maintain instruction sensitivity, and a flow-matching action decoder.

Experiments show state-of-the-art results on VLM benchmarks (ERQA, PhysBench, MME, MMMU, OCRBench, RealWorldQA, TextVQA) and VLA benchmarks (SimplerEnv-WidowX, SimplerEnv-GoogleRobot, LIBERO, RoboCasa-GR1), with notable out-of-domain gains.

Real-world Franka experiments show a 16.2 percentage point average improvement over π0.5 on single-object grasping and 14.0 points on long-horizon tasks.

The approach demonstrates that human-derived physical priors can improve downstream robot adaptation with limited robot data, though limitations remain in annotation quality, depth estimation errors, and embodiment differences.

02 From the paper

Abstract

Vision-language-action models have advanced rapidly, but robot trajectories alone provide limited coverage for learning broad physical understanding. PhysBrain 1.0 studies a complementary route: converting large-scale human egocentric video into structured physical commonsense supervision before robot adaptation. Our data engine extracts scene elements, spatial dynamics, action execution, and depth-aware relations, then turns them into question-answer supervision for training PhysBrain VLMs. The resulting physical priors are further transferred to VLA policies through a capability-preserving and language-sensitive adaptation design. Across multimodal QA benchmarks and embodied control benchmarks, including ERQA, PhysBench, SimplerEnv-WidowX, LIBERO, and RoboCasa, PhysBrain 1.0 achieves SOTA results and shows especially strong out-of-domain performance on SimplerEnv. These results suggest that scaling physical commonsense from human interaction video can provide an effective bridge from multimodal understanding to robot action.