VLA Finetuning Study: π0.5 Needs 2-5x Fewer Demos Than π0
rahulshivchand · x · 2026-11-11
A new paper systematically measures the sample efficiency of six VLA models (π0, π0.5, GR00T N1-N1.7) across four LIBERO suites.
Key findings
- Full-data finetuning hides big differences: models within 2 points on full LIBERO-Spatial differ by up to 3x in success rate with only 50 demonstrations
- π0.5 is the most sample-efficient model at every budget, needing 2-5x fewer demonstrations than π0, with the largest gap on long-horizon tasks
- Within GR00T, N1.5 adapts more efficiently than the stronger zero-shot N1.6/N1.7 on 3 of 4 suites
- Freezing the VLM during finetuning severely degrades the π models (whose VLM is jointly trained with the action expert) but barely affects GR00T
The authors argue sample efficiency deserves as much attention as zero-shot benchmark scores, since fewer demos mean cheaper data collection and faster adaptation.
More from Embodied
- Berkeley's Social Humanoid teaches Unitree G1 socially responsive interaction — berkeley_ai · 2026-11-11
- Robotics is not solved: frontier models are ~6 orders of magnitude off real assembly lines — ZeeshanZiaML · 2026-11-11
- MacBooks do have a hinge angle sensor—verified with Codex, prompt included — lipeng0820 · 2026-11-11
- DynaRobotics robot demo at CoRL sparks 'I want one at home' reactions — JasonMa2020 · 2026-11-11
- Humanoid Factory Deploy Stalls: Teleop Intervention Every 6 Minutes After 15 Weeks — ZeeshanZiaML · 2026-11-11
- CoRL 2026's wildest moments: 'Why do you care about delay?' and 'haptic doesn't matter' — breadli428 · 2026-11-11