Two months of π0.5 finetuning hits 98% success: data quality beats scaling
DominiqueCAPaul · x · 2026-09-24
A team spent two months ablating π0.5 finetunes on a real manufacturing task, reaching 98% policy success, and is publishing all results, data, and runs including failures. Key findings: 5x more data only moved success from 63% to 76%, making pure scaling the weakest lever; diversity beat quantity — 4 hours spread across five scenes beat 4 hours in the eval scene by 30 points; quality did best, with one clean, slower hour on top of 21 hours lifting 76% to 90%, more than the prior 17 hours. A quality-only recipe of 1h clean data plus 240 rollouts with human interventions (1.7h total) went from 28% to 88%. A commenter notes that from a UMI perspective scaling and diversity are easiest while quality is harder and interventions impossible — the exact opposite ordering — and accepts the challenge.
More from Embodied
- Baidu World 2026 to showcase Apollo Go robotaxi rides and Apollo World physical AI model — Baidu_Inc · 2026-11-12
- NVIDIA's open-source Alpamayo 2 Edge VLA driving model lands in Autoware — drmapavone · 2026-11-12
- A closer look at Astra's robot gripper design (and a joke at its expense) — _Stocko_ · 2026-11-12
- Robot startup passes first presale milestone with 10+ units claimed — R4nsack · 2026-11-12
- Runway launches Praxis-1, a 27B MoE world action model, with a 73ms Fast variant — ChongZzZhang · 2026-11-12
- Dream-1 on bimanual YAM robot folds a towel on voice command at CoRL — DJiafei · 2026-11-12