Recovery skills lift real-robot task success from 77.5% to 87.5% in Recova
qinzytech · x · 2026-10-07
Recova trains robot recovery skills separately from task execution: agents develop corrections in simulation, invoke them on hardware, and verify the scene before resuming. Adding recovery skills raised mean success from 77.5% to 87.5% across four real-robot tasks, with +15 points each on ring stacking and pencil-box packing where misplacements block remaining steps. Human demos of unresolved failures supply extra training data (20 trials per task/config). The takeaway: teams should collect demonstrations of scene repair alongside successful task runs.
More from Embodied
- Scoble visits a quick-turn Berkeley machine shop that robot prototype builders rely on — Scobleizer · 2026-10-07
- GroundedSLAM debuts, decisively beating all methods on Meta's egocentric SLAM benchmark — Scobleizer · 2026-10-07
- QF3 arrives: fast flow RL trains humanoids from scratch and fine-tunes VLAs — ChongZzZhang · 2026-10-07
- The ceiling for ambient, audio-only compute is extremely high: collapsing a whole workday into one thread — curious_vii · 2026-10-07
- MIT CSAIL Launches Ascent Lab, a Browser Platform for Robot Learning, With Solana Token — MIT_CSAIL · 2026-10-07
- Tesla's Halloween Mode Rolls Out: Ghost Costumes, Trick-or-Treat Sounds and Remote Scares — yunta_tsai · 2026-10-07