Dex-One2Many: one human video trains a dexterous hand that transfers zero-shot to real robots
furongh · x · 2026-10-09
- Dex-One2Many introduces a Real2Sim2Real pipeline that trains a dexterous-hand policy from a single human video.
- The key is neuro-symbolic abstraction: the system extracts the task's relational structure into "scene graphs" that encode stage-wise constraints (object positions, goal poses, grasps) as an inductive bias.
- These graphs generate diverse training states and RL rewards in simulation — constraints specify what must hold, RL discovers how to achieve it across configurations.
- One demonstration thus yields many valid ways to complete the task, and the learned policy transfers zero-shot to a real robot hand, generalizing to setups the video never showed.
More from Embodied
- Princeton's Intro to Robotics planning & control module: videos, notes, code all free — Majumdar_Ani · 2026-10-09
- Robotics IPOs in Hong Kong: 'still better than Solana memecoins' — LexSokolin · 2026-10-09
- China's AI-powered AR glasses: a real-world copilot guiding repairs and assembly — alexmacgregor__ · 2026-10-09
- 10+ robot teams compete to fold paper planes, testing tactile sensing and bimanual control — micoolcho · 2026-10-09
- Sabi raises $50M to build the world's most wearable BCI cap, backed by ex-OpenAI CPO Kevin Weil — LexSokolin · 2026-10-09
- Earth Rover Challenge wraps at IROS 2026: 12 teams, 4 tracks, 3 days of sidewalk robot navigation — micoolcho · 2026-10-09