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.
More from Embodied
- Tesla bulls watch Cybercab and Optimus buildout as capex and cash flow loom — JOBhakdi · 2026-07-22
- Tesla reportedly ships an HW3 FSD update distilled from the larger HW4 model — Teknium · 2026-07-22
- After IKFast, Rosen Diankov launches ssik as a new inverse-kinematics chapter — ZeYanjie · 2026-07-22
- Local Quantum LLM: single-photon QRNG drives token sampling across the multiverse — Reddactor · 2026-07-22
- Kinetix AI Unveils Full-Stack Embodied AI: 115-DOF Humanoid Plays Table Tennis, Dexterous Hand Goes on Sale — 新智元 · 2026-07-22
- NVIDIA Open Sources GPU-Accelerated Medical Physics Framework, Cutting Robot Training to 2 Minutes — NVIDIA Blog · 2026-07-22