Patch Policy Outperforms 7B VLA with 0.7% Parameters

Researchers introduced Patch Policy, a robotic manipulation method demonstrating that fine-grained control tasks do not necessarily require 7B-parameter Vision Language Models (VLMs). Using a fraction of the parameters, it achieved manipulation performance 18% higher than OpenVLA-OFT on a single 5090 GPU, drawing significant attention for its practical value.

Confirmed

The core architectural idea of Patch Policy is to preserve dense spatial visual features. It uses a ViT pretrained on internet data and completely frozen. Instead of extracting a single global vector or CLS token, it feeds all dense patch tokens directly into a small Transformer-based policy head for decision-making. According to author @furongh, this policy head also incorporates a block-causal mask mechanism during processing.

Why it matters

Traditional visuo-motor policies often face a dilemma: either compress the entire frame into a vector and lose fine spatial information, or train the visual backbone from scratch, thereby abandoning the benefits of large-scale pretraining. Patch Policy solves this problem with a minimal architectural change. Research data shows that with a parameter count of only 0.7% of OpenVLA-OFT, it significantly surpasses the performance of large VLA models, providing a new paradigm for highly efficient, low-compute robotic control.

2026-07-22 ~ 2026-07-23 · 5 related posts

• [source] Robotics paper says dense patch features beat bigger vision-language models — stepjamUK · 2026-07-22
• Patch Policy preserves dense spatial detail for robot manipulation — chris_j_paxton · 2026-07-23
• Patch Policy beats a fine-tuned 7B VLA by 18% with 0.7% of the parameters — ylecun · 2026-07-23
• Patch Policy beats OpenVLA-OFT with 0.7% of the parameters on one RTX 5090 — ChongZzZhang · 2026-07-23
• [source] Patch Policy uses a frozen internet-pretrained ViT and a block-causal mask for robot control — furongh · 2026-07-23