Baiz Tech's Abductive World Model Cuts Robot Training Data to 1/7 of Baselines
机器之心 · wechat · 2026-10-11
Baiz Tech, a Hefei startup less than six months old founded by a USTC team, unveiled a three-layer embodied brain: the Abductive World Model (AWM), edge inference engine BAIZ-RUN, and a compute terminal. AWM 'predicts, then abduces'—decomposing the world into entities, dynamics, and relations—and learns from unlabeled video. On Push-T it hits 75% success with just 2,000 trajectories (baseline Causal-JEPA needs 14,728); vs. V-JEPA2 it improves action-understanding Top-1 accuracy by 68%. The paper and code are public.
BAIZ-RUN uses error-aware quantization (2-bit weights, 4-bit activations), cutting memory 2.56x and boosting inference frequency 94.6% with under 1% accuracy loss on AMD edge GPUs, with multi-backend support for NVIDIA/AMD/Ascend—aiming to be the vLLM of robotics. Three industrial use cases (material loading, inspection, patrol) are signed and validated, with partners including Wanyu, Weigang, ZERITH, and LimX Dynamics; the company raised tens of millions of RMB within three months, valuing it above 400M RMB.
More from Embodied
- Glovebox launches: open-source tool runs Half-Life: Alyx standalone on Vision Pro — banteg · 2026-10-11
- Open converter lets you port custom sim environments to MuJoCo and Isaac Lab — joemeno · 2026-10-11
- Musk Slams Rivals' Robot Demos as Fake — Critics Point to Teleoperated Optimus — flowersslop · 2026-10-11
- Bezos criticizes autonomous car ride smoothness; Andrew Chen fires back — andrewchen · 2026-10-11
- Nils Pihl to keynote Humanoid Hub Conference on the missing shared context layer for robots — broodsugar · 2026-10-11
- NUS MAGIC Lab hiring: postdocs, PhD students and robotics researchers in Singapore — DJiafei · 2026-10-11