[HUGGINGFACE]score: 0.62

Object-Centric Residual RL Transfers Sim-Trained Policies to Real VLAs Zero-Shot

June 16, 2026

A residual RL framework trained purely in simulation refines frozen VLA actions using object poses rather than raw images or privileged simulator state, sidestepping the visual domain gap and avoiding costly real-world RL. The compact object-centric observation space enables zero-shot sim-to-real transfer on top of existing VLAs without retraining them.

HOW THIS AFFECTS YOU

●

builderYou can potentially bolt this residual RL layer onto a frozen VLA to improve manipulation precision without collecting real-world training data or retraining the base model.

●

researcherWorth watching because the object-pose observation space is a concrete architectural choice that avoids the three standard failure modes of sim-to-real residual RL — privileged state, visual gap, and real-world cost.

read original ↗huggingface.co

← back to feed