DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43%
liuziwei7 · x · 2026-10-27
DONIC from Ziwei Liu's team at Ropedia trains a single generalist policy for two-handed dexterous reference tracking, where human demos disagree on which hand does what. It mixes PPO with behavior cloning so successful rollouts become new demos, uses per-hand contact-role rewards and clip-centered advantages. With the BiHOI dataset (1,575 sequences, 229 objects, 14.77h), it reaches 86.79% success on unseen trajectories vs 9.43% for the best prior tracker (Dexplore-bi), and 57.32% vs 11.59% on held-out objects. Caveat: still simulated in Isaac Gym on single-object tasks, with real-world eval clips shown.
More from Embodied
- Apple responds after iPhone Duo inner displays crack along the hinge — mobileraj · 2026-10-27
- VLA models encode task concepts as curved multi-dimensional geometry, 19-author study finds — ninamiolane · 2026-10-27
- Munro Live Posts Full Cybercab Engineering Deep-Dive Video — yunta_tsai · 2026-10-27
- Tesla's Robotaxi Fleet Keeps Growing in Austin — yunta_tsai · 2026-10-27
- Building Halo Infinite Forge maps with local LLMs, DCjson and Blender MCP — swagonflyyyy · 2026-10-27
- SoftBank-Backed XAG Ships Drones With Autonomous Docks That Clean, Recharge and Refill — pstAsiatech · 2026-10-27