MDL-9017EST.2025 · IDX.727
Robotics AIIn production

AlphaBrain (GOVLA)

AI² Robotics (Zhipingfang) · China · 2025

The vision-language-action brain behind AI² Robotics' AlphaBot 2 — one model for seeing, reasoning and moving the whole body.

wujec.ai score

7.0/10

Community score

no votes yet
Sign in to rate

AlphaBrain is the software layer that AI² Robotics sells alongside its AlphaBot robots, built around an architecture the company calls GOVLA — Global and Omni-body Vision-Language-Action. The pitch is that one trained model replaces the usual stack of separate perception, planning and control modules: it takes in what the robot sees and what it is told, reasons about a multi-step job, and drives every joint in the body rather than an arm alone. The company describes a two-speed design. A fast loop produces motion at control rates, while a slower reasoning process plans the task and periodically re-conditions the fast loop — an arrangement that lets a robot keep moving smoothly while it is still working out what to do next. AI² Robotics says inference runs on the robot itself, which is what allowed a public demonstration at the BRIDGE Summit in December 2025 to survive poor lighting, crowds and an unreliable network. On the same hardware and with no task-specific scripting, AlphaBot 2 was shown making coffee and playing drums. AI² Robotics publishes no parameter count, training-data description or benchmark table for the commercial model, and the developer platform requires an account, so the figures below are deliberately sparse. The closest public technical reference is Fast-in-Slow (FiS-VLA), a NeurIPS 2025 paper co-authored by the company's chief executive Yandong Guo, which describes the same fast-inside-slow idea in detail: a LLaMA-2 7B backbone with SigLIP and DINOv2 vision encoders, a 3D point-cloud tokenizer, 117.7 Hz control and success rates 8 percent above the prior state of the art in simulation and 11 percent in real tasks. The paper does not claim to be AlphaBrain, and wujec.ai does not treat it as the product's specification — only as the published work closest to it.

#vision-language-action#embodied AI#whole-body control#on-device#China
Official website

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review