Patch Policy: Boosting Embodied Control via Dense Visual Representations

NielsRogge · x · 2026-08-13

The post introduces a robot learning architecture named Patch Policy. Current robot policies either compress observations into a single global token or rely on heavy Vision-Language Models (VLMs), sacrificing fine-grained details and speed.\n\nPatch Policy uses a minimal architectural extension to let transformer policies consume dense pre-trained patch tokens directly. Its core is a block-causal attention mask that preserves temporal causality while processing many patch tokens.\n\nExperiments show a 40% relative improvement over policies using global-pooled representations. It also surpasses fine-tuned OpenVLA-OFT while using only about 0.7% of its parameters.

Original post →

More from Embodied

Embodied channel →