UBTech Showcases Factory Line: Real-World Data Dominates Training
量子位 · wechat · 2026-08-22
On-Site: Factory Replication
At WRC, UBTech (优必选) replicated client factory lines 1:1. Industrial humanoid robots Cruzr S2/Y1 performed continuous, unattended tasks like automotive part handling and logistics sorting with <1mm precision and a rate of 1100 items/hour, capable of autonomous error correction.
Embodied AI Architecture
UBTech revealed its "Embodied Brain" system:
- Thinker Foundation Model: Handles visual, language, and spatial understanding (up to 100B parameters).
- Thinker-WM World Model: Predicts action outcomes and physical laws; topped the LIBERO benchmark.
- Thinker-VLA: Manages real-time end-side execution for trajectory and force control.
Edge Deployment Optimization
To reduce latency and cost, models are pushed to the edge. Inference efficiency improved by 176%, and storage usage dropped by 60%, allowing full operation on low-power ARM boards.
Data Loop & Commercialization
- Data Mix: Real-world robot data accounts for 60%-70%, Ego data 20%, and simulation only 10%.
- Financials: 2025 revenue exceeded 2 billion RMB with 13,838 humanoids sold. 1,079 were full-sized industrial robots, with >80% deployed in strict sectors like automotive and 3C manufacturing.
- Ecosystem: Joint venture with Maxio for edge chips, acquisition of Fenglong for components, and a 10,000-unit capacity smart factory with Siemens.
Related event: UBTECH Showcases Real-World Robot Work at World Robot Conference(2 posts)→
More from Embodied
- Humanoid Robots Outrun Humans in Race: X-humanoid Wins — CyberRobooo · 2026-08-22
- Opinion: AI hardware should sell taking your context with you, not locking it in — op7418 · 2026-08-22
- Humanoid hardware surpasses humans; system integration remains the gap — ChongZzZhang · 2026-08-22
- Chinese Team Launches DeepSoma Platform to Crack PhysicalAI Challenges — 量子位 · 2026-08-22
- Octopus Limb Control Architecture Revealed, Inspiring Soft Robotics — MacrinePhD · 2026-08-22
- View: Robotic Models to Be Trained by LLMs via Real2Sim RL — YouJiacheng · 2026-08-22