Zero-HLM trains humanoid loco-manipulation with zero real robot data, lifting success 23% to 69%
siyuanhuang95 · x · 2026-10-29
Researchers unveil Zero-HLM: generalizable humanoid whole-body loco-manipulation without any real robot demonstrations.
The core recipe is real2sim2real:
- 📱 Capture real scenes with a smartphone
- 🌐 Build visually faithful, physics-ready digital twins with Astra
- 🦾 Generate whole-body demonstrations in simulation
- 🚀 Transfer policies zero-shot to a real humanoid
Simulation-only training transfers zero-shot to a Unitree G1 across three long-horizon tasks; pooled real-world success across five generalization settings rises from 23% (real-data training) to 69% (their simulation data), questioning how much real robot data is actually needed.
More from Embodied
- StackChan desk robot gains meme generation, email, and a rave mode — alexcovo_eth · 2026-10-29
- Michigan's Open Source Leg: a modular open-source powered prosthetic leg — lukas_m_ziegler · 2026-10-29
- Espressif launches ESP-Mosaico: an ESP32-S31 kit built for vibe coding — songguoxiansen · 2026-10-29
- Robot 'Made in USA' Tariff Rules: 65% Domestic BOM Plus Substantial Transformation — MatthewChang · 2026-10-29
- Matrix-3 humanoid debuts at Wuzhen World Internet Conference, built for service work — CyberRobooo · 2026-10-29
- China's ultra-realistic androids blur the line between human and machine — LexiLove · 2026-10-29