π0.5 study: raising demos from 10 to 438 lifts Place Ball in Bowl success from 7% to 83%
stingning · x · 2026-10-27
The author shares single-task π0.5 experiments on how demonstration count affects policy success:
- Place Ball in Bowl: success rises from 7% to 83% as training demonstrations grow from 10 to 438.
- Squeeze Scissors: success rises from 1% to 44% as demonstrations grow from 10 to 439; the functional scissors task stays harder.
Success increases at every tested dataset size. Runs use fixed training updates and a single seed per run, so the authors frame this as evidence that more demonstrations help on these two tasks, leaving broader scaling behavior open.
The same thread lists three recurring failure patterns in trained-policy rollouts:
- Premature activation: in Move and Click Mouse, early clicks terminate 443 of 800 episodes, often while the hand is still acquiring or positioning the mouse.
- Coordination: in Set Dual Knobs, both knobs reach targets simultaneously in only 38 of 800 episodes, and just 5 complete the full stability-and-release requirement.
- Goal re… (truncated in source)
More from Embodied
- UniAction: unified action conditioning lets one world model handle mixed action types — ZeYanjie · 2026-10-27
- StackChan robot performs an AI-generated skit, orchestrated by Hermes Agent via JSON script — alexcovo_eth · 2026-10-27
- DONIC hits 86.79% on bimanual dexterous tracking, crushing prior SOTA's 9.43% — liuziwei7 · 2026-10-27
- Bedrock Robotics raises $600M as its fully autonomous excavators do paid work on live job sites — Sethwinterroth · 2026-10-27
- AW-0 debuts: one generative world action model plans in code, predicts, and controls robots — trevordarrell · 2026-10-27
- Skyfall AI's SKY-001 world model turns sketches into CAD builds in 3.9s, vs 199s for GPT-6 Astra — rohanpaul_ai · 2026-10-27