DeepRobotics (Yunshenchu) has released five research results around its self-developed embodied brain, DEEPRobotics-Brain, sketching a complete architecture across the full embodied-intelligence chain.
The world model DR-WM-1.0 is the first in quadruped world models to embed physical properties such as friction, roughness and stiffness, so predicted futures are not only visually realistic but physically plausible — letting a robot 'pre-run' a slippery shortcut versus a detour before deciding. The embodied base model DR-VLA-2.0 learns human social and traffic rules from 10,000 hours of first-person human video plus 2,000 robot demonstrations, improving human-like interaction.
The suite also includes the embodied agent DR-Agent, the policy-optimization framework DR-APO and the data-closed-loop platform DR-Benchmark. The work was co-developed with Zhejiang University, Beijing Institute of Technology, Tsinghua University, MIT and the University of Illinois Urbana-Champaign.
Source: NetEase / Sina Finance (paraphrased and translated from Chinese media; reprinted for informational purposes with attribution).
