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.
Related event: Patch Policy Outperforms 7B VLA with 0.7% Parameters(5 posts)→
More from Embodied
- Wayve CEO argues end-to-end deep learning should power driving, safety and world models — alexgkendall · 2026-10-27
- Mundo turns a few room photos into an explorable, simulatable 3D world — at 1.5 FPS — mark_k · 2026-10-27
- Zurich Physical AI Hackathon puts attendees in a room with $1.5M of humanoids — ___Mufasaa · 2026-10-27
- Apple responds after iPhone Duo inner displays crack along the hinge — mobileraj · 2026-10-27
- VLA models encode task concepts as curved multi-dimensional geometry, 19-author study finds — ninamiolane · 2026-10-27
- Munro Live Posts Full Cybercab Engineering Deep-Dive Video — yunta_tsai · 2026-10-27