Fine-tuning π0.5 on a real factory task: data diversity beats raw scale, hitting 98% success
DominiqueCAPaul · x · 2026-09-24
Dream Machines' first blog post documents two months of ablation experiments fine-tuning Physical Intelligence's open-source VLA model π0.5 on a real manufacturing task from a German manufacturer, reaching 98% policy success rate. All results, data, and runs—including failures—are being published.
- Pure scaling was the weakest lever: 5x more data only improved success from 63% to 76%.
- Diversity got more from the same hours: 4 hours of data spread across five scenes beat 4 hours in a single scene.
The task requires two arms to transfer actuators from a box into a fixture with correct orientation. The authors note that neither the π0.5 paper nor the official repo offers fine-tuning recommendations, and no other company has published results, so these findings aim to save others time on tuning and data collection.
More from Embodied
- Humanoid robot soccer demo: 140 simulated years of self-play training — deepakpathak · 2026-09-24
- The world is the installed base: the underrated case for humanoid robots — r0ck3t23 · 2026-09-24
- GTSAM 4.3 ships with legged navigation, CUDA-accelerated factor graphs and continuous-time trajectory estimation — fdellaert · 2026-09-24
- Workbench open-sources an agentic lab workspace for hardware engineering on the Ambion kernel — andreisavu · 2026-09-24
- Hugging Face Tutorial: Accelerating Robotics Simulation with NVIDIA Warp and MjWarp — Hugging Face Blog · 2026-09-24
- Skild AI trains Unitree G1 to play soccer with 140 years of simulated self-play — Distinct-Question-16 · 2026-09-24