Robotics paper says dense patch features beat bigger vision-language models

stepjamUK · x · 2026-07-22

A robotics paper argues that precise manipulation does not need a billion-parameter vision-language model if the policy can preserve dense spatial features.

What Patch Policy changes

It avoids pooling observations into a single CLS/global vector.
It also avoids training a new visual backbone from scratch.
Instead, it feeds frozen dense patch features from a pretrained ViT directly into a transformer policy head.

Why it helps

The method keeps spatial attention within each frame bidirectional, while using causal masking across frames so temporal order stays intact.

Reported results

Across four simulated and three real-world suites, dense patch features beat global-pooled baselines by 40% relative.
Compared with a fine-tuned OpenVLA-OFT, Patch Policy is 18% better with about 0.7% of the parameters and 11 ms inference.
Replacing patch features with pooled or convolutionally compressed ones hurts performance across the board.

The author’s takeaway: for manipulation, precision comes from preserving patch-level structure, not from making the backbone bigger.

Original post →

Robotics paper says dense patch features beat bigger vision-language models

More from Embodied