Dex-One2Many learns dexterous manipulation from one human video, +71% on unseen setups
Jusuk Lee · hf · 2026-10-12
Researchers from Seoul National University, University of Maryland, and KAIST present Dex-One2Many, a real-to-sim-to-real framework that learns a generalizable dexterous manipulation policy from a single human video.
- Imitation-based methods overfit to demonstrated motions, while pure RL struggles with exploration in multi-stage tasks
- Key idea: abstract the video into sequential scene graphs that guide RL — the graphs act as generative constraints for sampling diverse reset states and provide dense per-stage rewards
- Because graphs constrain relations rather than exact poses, reset states cover object poses and grasps beyond the video while keeping exploration short and guided
- Trained entirely in simulation, the policy transfers zero-shot to a real multi-fingered hand
- Across five tool-use tasks, it beats baselines by 6.5% on seen configurations and by 71% in unseen scenarios
More from Embodied
- ENCORE: Coding Agents Write Robot Policies From 5 Demos, Hitting 96.3% on LIBERO-PRO — 2077AIDataFoundation · 2026-10-12
- ARTISAN: Multi-Agent Pipeline Generates Physically Grounded Articulated 3D Assets with New Benchmark — elliottszwu · 2026-10-12
- $84,000 Boston Dynamics robot washes a lioness, sparks animal-intervention debate — CurieuxExplorer · 2026-10-12
- San Francisco robot fight club levels up: 200-pound humanoids duel with steel blades — cixliv · 2026-10-12
- 1X to supply up to 10,000 NEO humanoid robots to EQT portfolio companies — CyberRobooo · 2026-10-12
- ByteDance's unreleased world model demoed: Stephen Chow film license, trajectory-to-previs — vista8 · 2026-10-12