Proxy Policy Steering adapts frozen VLA models to new tasks at inference time
weichiuma · x · 2026-09-14
Researchers propose Proxy Policy Steering (PPS), which specializes VLA (vision-language-action) models to new tasks at inference time without touching base model weights.
- Instead of direct fine-tuning, it trains two lightweight proxy policies
- Their velocity-space residual steers the frozen base model's behavior for the new task
The upshot: large VLA bases stay frozen while cheap external modules handle task adaptation, cutting customization cost.
More from Embodied
- How to judge robot demos: change the environment, then count the failures — paigeinsf · 2026-09-14
- MicroDuck robot learns to climb, and walking was just the start — econoar · 2026-09-14
- Tesla China to exhibit Cybercab in Shanghai and Beijing from September 17 — yunta_tsai · 2026-09-14
- MIT's HardFlow enforces hard constraints on generative AI without retraining — MIT News AI · 2026-09-14
- MIT spinout Atlas turns plastic waste into building materials with AI robotics — MIT News AI · 2026-09-14
- UBTech Opens World's First 10,000-Unit Humanoid Robot Factory; SK's AI Data Center Hits 900MW — 创业邦 · 2026-09-14