Cornell's Proxy Policy Steering Adapts Frozen Robot Policies Without Touching Weights

philfung · x · 2026-09-28

A Cornell team (including Weichiu Ma and Kuan Fang) proposes Proxy Policy Steering (PPS), an inference-time method that adapts a large frozen robot policy to new tasks from few demonstrations without modifying its weights.

It trains two small networks: a reference proxy distilled from the frozen base policy, and a task proxy fine-tuned from that reference on task demonstrations. At sampling time it steers the base model's diffusion trajectory with a velocity-space residual: vPPS = vbase + γ(vtask − vref).

Notably, the base model's parameters are never modified and aren't even required. It is evaluated on eight real-world manipulation tasks (coffee brewing, flower insertion, jeans folding, shoe retrieval, tissue wiping, etc.) and four simulation tasks.

Original post →

More from Embodied

Embodied channel →