Freezing the VLM costs π models 23-56 points but barely hurts GR00T
rahulshivchand · x · 2026-11-11
Thread takeaway: VLAs are usually fine-tuned on downstream tasks, so sample efficiency matters as much as zero-shot performance. Ablations show freezing the VLM backbone during fine-tuning drops π models' success by 23-56 points while GR00T changes by at most 6 points, pointing to large pretraining recipe differences.
More from Embodied
- Maestro paper: cheap Gemini 2.5 beats π₀.5 baselines at zero-shot robot control — dineshjayaraman · 2026-11-11
- OpenDDE: open-source all-atom drug discovery engine hits 51% on hard antibody-antigen benchmark — teortaxesTex · 2026-11-11
- Tesla demos Optimus walking inside a world model at CoRL, real policy in control — Scobleizer · 2026-11-11
- NUS Proposes WM-PPO: World-Model-Predicted Successor States Improve PPO Policy Learning — NationalUniversityofSingapore · 2026-11-11
- DexForesight Boosts Dexterous Manipulation Success From 49.4% to 60.5% With Implicit Spatial Foresight — Sharpa-Robotics · 2026-11-11
- NTU Study Finds Time-Indexed Memory Makes VLA Robots Mistake Pauses for Progress — NanyangTechnologicalUniversity · 2026-11-11