VLS: training-free VLM steering of frozen robot policies wins CoRL and CVPR workshop awards

micoolcho · x · 2026-09-29

RoboPapers teased a podcast on VLS (Vision-Language Steering), a CoRL 2026 paper from UW, Oxford, NUS and Allen AI that also won Best Paper Runner-Up and Outstanding Paper awards at CVPR 2026 workshops. VLS is a training-free framework that adapts frozen diffusion/flow-matching robot policies at inference time: a VLM generates differentiable rewards for partially denoised action proposals, steering sampling so policies handle out-of-distribution objects, scenes, and instructions without fine-tuning. Code and arXiv are public.

Original post →

More from Embodied

Embodied channel →