Former Xiaomi EV L3 Lead Launches Embodied AI Startup for Robotic Brains
Wang Naiyan, former L3 autonomous driving technology lead at Xiaomi EV, has launched an embodied AI startup seeking funding to build highly generalizable robotic brains. His core R&D team consists of over a dozen members, including several former colleagues from TuSimple. This news highlights an accelerating industry trend of top autonomous driving executives moving into the robotics sector to adapt vehicle automation techniques for physical AI. Applying proven L3 autonomous driving logic to robotics could significantly advance how embodied agents adapt to unstructured real-world environments. The startup's technical approach relies on applying reinforcement learning during the pre-training stage of action models using domain randomization. By setting foundational first-principles reward functions like safety boundaries and physical laws, the model learns through millions of self-play iterations to achieve stronger generalization.
## BACKGROUND
Embodied AI connects artificial intelligence with physical robotic systems, enabling machines to perceive, reason, and act in physical environments rather than operating purely in digital spaces. Domain randomization is a technique in reinforcement learning that intentionally randomizes simulation parameters during training, helping AI models remain robust when deployed in unpredictable real-world scenarios.