π0.5 fine-tuning ablations: hard lighting halves expert checkpoint success, full fine-tune unaffected
philfung · x · 2026-10-20
PrismaX published an openpi/π0.5 fine-tuning how-to with ablation tests behind their IROS demo. Key finding: hard lighting changes cut the expert-only checkpoint's success rate roughly in half, while the full fine-tune was completely unaffected — suggesting full fine-tuning is far more robust to visual distribution shift. The article walks through training-setting ablations (data blends are proprietary), making it a rare practical reference for fine-tuning π0.5-style robot policies.
More from Embodied
- "Robots are going underground, figuring out their own routes" — a look past chatbot debates — VraserX · 2026-10-21
- Deepak Pathak: There Is No Perfect Data Source for Robotics — chris_j_paxton · 2026-10-21
- Discrete action representations match or beat continuous ones for robot policies — GuanyaShi · 2026-10-21
- Developer gave his Grok bot $1,000 to build itself a body — parts keep arriving — Baconbrix · 2026-10-21
- Tuesday Labs opens $200 reservations for 999 numbered Voxel Series-1 home robots — testingcatalog · 2026-10-21
- Tuesday Labs' home robot Voxel charms with Pixar-collaborated personality design — cneuralnetwork · 2026-10-20