π0.5 Fine-Tuning Study: 1 Hour of Clean Data Beats 17 Hours of Scale
DominiqueCAPaul of the Dream Machines team published a first blog post on September 24, sharing the complete results of a two-month finetuning ablation study on Physical Intelligence's open-source VLA model π0.5, run on real manufacturing tasks at a German manufacturer. Policy success rate ultimately reached 98%, with a commitment to release all results, data, and experiment logs (including failed runs).
Confirmed
- Simply scaling data is the weakest lever: a 5x increase in data volume only lifted success rate from 63% to 76%.
- Data diversity pays off more: 4 hours of data spread across 5 training scenarios beats training only on the evaluation scenario.
- 1.7 hours of high-quality data with interventions outperformed 21 hours of pure demonstration data; the author sums it up as "1 hour of clean data beats our previous 17 hours."
- All data and failed runs have been open-sourced.
Why it matters
This is a rare public practice of systematically ablating an open-source VLA model on real production-line tasks. The results show that the bottleneck in finetuning robot policies lies not in data volume but in data quality and diversity, and that data labeled with human interventions delivers far better value than pure teleoperation demonstrations — directly relevant to industrial deployment. Full open-sourcing (including failed experiments) also gives the community a reusable baseline.
2026-09-24 ~ 2026-09-24 · 6 related posts
Primary sources
- [source] π0.5 finetuning on real manufacturing: 1 hour of clean data beat the previous 17 — DominiqueCAPaul · 2026-09-24
- [source] Fine-tuning π0.5 on a real factory task: data diversity beats raw scale, hitting 98% success — DominiqueCAPaul · 2026-09-24
- 1.7h of intervention data beats 21h of demos: finetuning π0.5 on manufacturing — DominiqueCAPaul · 2026-09-24
- Two months of π0.5 finetune ablations: 1 hour of clean data beats 17 hours of scale — DominiqueCAPaul · 2026-09-24
2 near-duplicate retellings: DominiqueCAPaul · DominiqueCAPaul