π0.5 finetuning ablations: 1 clean hour beats 17 hours of raw data scaling
DominiqueCAPaul · x · 2026-09-24
Two months of ablations on π0.5 finetunes for a real manufacturing task, reaching 98% success with all results, data and runs published. Key findings: 5x data only lifted 63%→76%; spreading 4h across five scenes beat 4h in the eval scene by 30pp; one clean, slower hour on top of 21h took 76%→90%; and just 1.7h total (1h clean data + 240 human-intervention rollouts) jumped 28%→88%. Quality and diversity beat pure scale.
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