Two months of π0.5 finetune ablations: 1 hour of clean data beats 17 hours of scale
DominiqueCAPaul · x · 2026-09-24
The author spent two months ablating π0.5 finetunes on a real manufacturing task, reaching a 98% policy success rate, and will publish all results, data, and runs. Key findings: 5x more data was the weakest lever (63%→76%); spreading 4h across five scenes beat 4h in the eval scene by 30pp; adding just 1 clean hour on top of 21h jumped 76%→90% — more than the previous 17 hours; and 1h clean data plus 240 human-intervention rollouts (1.7h total) lifted 28%→88%. Data quality, not scale, is the lever.
More from Embodied
- UBTECH sells 400 U1 Pro humanoids to Yongda Auto in ~$10M deal — CyberRobooo · 2026-11-12
- 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