Ex-Zhiyuan PM's embodied data startup tops $3M in orders; bottleneck is data processing, not collection
快鲤鱼 · wechat · 2026-11-09
Song Zhihan, founder of embodied-data startup Xingyi Intelligence (and former humanoid product manager at Zhiyuan Robotics), discusses the industry with Chuangyebang.
Company & business
- Focuses on processing and high-precision annotation of first-person human data (hand/body pose, spatiotemporal consistency) rather than collection; a core team member built the gesture base model for ByteDance's Pico.
- Orders exceed 20M RMB; clients include embodied model company Being (whose H0.8 implicit tactile-visual model used Xingyi data) and JD.com (JD provides capture hardware and scenes, Xingyi handles processing/annotation).
Key takes
- Data accumulation won't be the bottleneck — processing will: fewer than 5% of embodied datasets on one public cloud platform are finely processed for reuse; recovering hand pose under occlusion, motion blur and low light is the hard part.
- Human data is cross-embodiment and teaches manipulation logic; future training may be 90–99% human data plus minimal real-robot data; simulation adds limited intelligence.
- Multimodal model companies currently demand first-person human data more than embodied labs — GPT-6 Astra reportedly trained on millions of hours of it.
- On Fei-Fei Li's World Labs selling to AMD: big tech entering world models signals accelerating route convergence; a reasonable阶段性 outcome for her.
- Data companies' moats: capture side competes on hardware/labor cost; processing side on efficiency and precision. Xingyi bets on a unified Transformer base predicting spatiotemporal pose consistency to scale capacity exponentially rather than linearly hiring.
More from Embodied
- RoboMoCap dataset released: 200+ hours of 120Hz whole-body mocap for humanoid robot learning — DJiafei · 2026-11-09
- Tesla cuts Optimus AI6 chip memory to 144GB, a 33% reduction amid DRAM shortage — JOBhakdi · 2026-11-09
- Tesla expands Austin Robotaxi zone to roughly 3x Waymo's coverage area — Polymarket · 2026-11-09
- TacClaw-Infinity: 10,000-Hour Vision-Tactile Dataset Lifts π0.5 Success from 50% to 74% — CyberRobooo · 2026-11-09
- Waymo explains at Hot Chips why camera-only autonomy plateaus at 99.9% — 4310sy · 2026-11-09
- 'The future is here, just unevenly distributed': humanoid robot spotted at a conference — technollama · 2026-11-09