~400 human corrections push robot espresso success past 90%, doubling throughput
micoolcho · x · 2026-10-25
Quoting @rahimmirani: most of the robot data everyone is fighting over is the wrong kind. Teleop gives clean demos of the easy 90%, but failures and recoveries — the hard part — go unlogged because operators just retry.
@physicalintelligence demonstrated the alternative with pi0.6 on a real espresso task:
- Took the model trained on normal demos and ran it on the robot
- Added 400 human corrections (interventions when it messed up) plus its own runs, failures included, and trained on all of it
- Result: throughput more than doubled, success went past 90%, and failure rate on the hardest tasks roughly halved
Same model, same robot — just a different kind of data. Failure-and-recovery data may be the real scarce resource in robot learning.
More from Embodied
- China Mobile demos brain-controlled humanoid robot with dexterous hand via non-invasive BCI — CyberRobooo · 2026-10-25
- iPhone Duo screen cracks cluster in same spot, pointing to hinge design flaw — oran_ge · 2026-10-25
- Four tiny Microduck open-source robots steal the show at NVIDIA GTC Berlin keynote — Roger_M_Taylor · 2026-10-25
- Robot IPO window closes: China's embodied AI firms stall after Unitree listing — 量子位 · 2026-10-25
- Beff Jezos wants AI glasses plus a subvocalization mic to talk to Grok away from his desk — beffjezos · 2026-10-25
- Getting your AI companion dot into the Halloween spirit — blakesamic · 2026-10-25