Fine-tuning π0.5 with recovery demos beats adding more success demos for robot reliability
k7agar · x · 2026-10-27
Researchers fine-tuned Physical Intelligence's π0.5 on two real-world manipulation tasks to study what happens when robots make mistakes.
The core question: as robots get better at completing tasks, can recovery demonstrations — data showing how to fix errors — improve policy reliability more effectively than simply adding more successful demonstrations? The thread walks through the comparison experiments on both tasks.
More from Embodied
- Terminal color themes become a physical button that pings you when agents finish — dee_hw · 2026-10-27
- Houmo LQ50-24 M.2 accelerator claims 160 TOPS but real-world tests are scarce — lazy-kozak · 2026-10-27
- Scoble: Tesla's new Texas plant targets 10M Optimus robots a year, market hasn't priced it in — Scobleizer · 2026-10-27
- Machenta Emerges From Stealth With $40M Seed for Instant Robot Learning — Ed__Johns · 2026-10-27
- Unitree Go2 Robot Dog Chat Module Had 9.4-CVSS Remote Command Injection Bug — AIFlow_ML · 2026-10-27
- ETH humanoid uses head-mounted LiDAR for onboard terrain mapping in 30ms — ChongZzZhang · 2026-10-27