UCLA's DexDirect cuts demo cost: 17.2x more demos than AnyTeleop, 90% policy success
DennisHongRobot · x · 2026-11-18
DexDirect, accepted to IEEE-RAS Humanoids 2026 from UCLA, is a low-setup interface for collecting dexterous manipulation demos: the operator drags a 6-DoF gravity-compensated arm while one webcam retargets their other hand onto a 16-joint dexterous hand.
Key results:
- 17.2x more successful demos than vision-based AnyTeleop, 3.2x vs TeleDex
- NASA-TLX workload of 2.01 vs 3.08/3.81 baselines, sharply lower mental demand
- A diffusion policy trained on its demos hits 90% success on cube pick-and-place
Code to be released; direct kinesthetic guidance plus hand retargeting looks like a scalable demo-collection recipe.
More from Embodied
- Tokyo Hosts Free 'Put AI Agents on Gadgets' Build Event on Dec 19 — alexcovo_eth · 2026-11-18
- Even bull riding is running on RL now, quips roboticist Kyle Morgenstein — KyleMorgenstein · 2026-11-18
- Genesis's first US hire looks back: an ETH-stacked Zurich office hidden in the Bay Area — theo_gervet · 2026-11-18
- "No sim2real gap if you only deploy in sim" — who is building this? — rahulshivchand · 2026-11-18
- Maker adds a new feature to the googly-eye clock and everyone wants one — alexcovo_eth · 2026-11-18
- Robo-ValueRL rates actions by estimated task progress to learn from messy robot data — burkov · 2026-11-18