Ex-OpenAI RLHF researcher's robotics startup ships open nav model LightNav-0 and resilient control LightREACT in 10 days

机器之心 · wechat · 2026-09-12

QijiXin reports on LiangYuan XinChuang, founded by ex-OpenAI RLHF researcher Jiang Xu, which applies the LLM playbook of pretraining-alignment-deployment to robotics. In early September it shipped two releases in ten days: LightNav-0, an open-source general navigation model built on Qwen3-VL-4B-Instruct that works zero-shot across humanoid, quadruped, wheeled, and flying robots with monocular RGB only—beating multi-camera systems on 10 public sim benchmarks, trained on 4,000+ hours of synthetic data from 2,000+ real scenes; and LightREACT, a single-Transformer resilient whole-body control policy that handles walking, fall recovery, limping/hopping with failed joints, and crawling, with no fault labels or mode switching—demos show a humanoid with tied legs freeing itself and standing back up. The team views resilience as the 'last mile' for deployment and data flywheels, and a consumer robot product is in development.

Original post →

More from Embodied

Embodied channel →