Two Months Fine-tuning π0.5 on Real Manufacturing Tasks: 1 Hour of Clean Data Beats 17 Hours of Scale

DominiqueCAPaul of the Dream Machines team published the team's first blog post on September 24, sharing complete results of a two-month fine-tuning ablation on Physical Intelligence's open-source VLA model π0.5 using real manufacturing tasks at German manufacturers: policy success rate ultimately reached 98%, with all results, data, and every experiment run (including failures) open-sourced.

Confirmed

• Pure data scaling is the weakest lever: 5x more data only improved success from 63% to 76%.
• Data diversity pays more: 4 hours of data spread across 5 training scenes outperformed training only on the evaluation scene.
• Human-intervention data far outperforms pure demos: 1.7 hours of high-quality intervention-labeled data beat 21 hours of pure demonstration data; under 1 hour of DAgger-style human intervention data produced a major jump in success rate — summarized by the author as "1 hour of clean data beats the previous 17 hours."
• All data and failed runs have been open-sourced.

Why it matters

This is a rare public, systematic ablation of an open-source VLA model on real production-line tasks. The results show the bottleneck in robot policy fine-tuning is data quality and diversity rather than volume, and that intervention-labeled data offers far better ROI than pure teleoperation demos — directly relevant to industrial deployment. Full open-sourcing, including failed experiments, gives the community a reusable baseline.

2026-09-24 ~ 2026-09-25 · 9 related posts

• [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
• Two months of π0.5 finetuning ablations: diversity beats raw scaling, 98% success — philfung · 2026-09-24
• Two months of π0.5 finetuning hits 98% success: data quality beats scaling — DominiqueCAPaul · 2026-09-24
• Fine-tuning π0.5 on a real manufacturing task: <1 hour of DAgger data drove big success-rate gains — DominiqueCAPaul · 2026-09-25

2 near-duplicate retellings: DominiqueCAPaul · DominiqueCAPaul