Patch Policy preserves dense spatial detail for robot manipulation
chris_j_paxton · x · 2026-07-23
Patch Policy keeps spatial detail for robot control
The paper argues that visuomotor policies often lose the fine-grained spatial information needed for manipulation. Its answer is to keep dense patch features from a ViT instead of pooling everything into a single vector or training a vision backbone from scratch.
Main idea
- Encode multi-view observations as dense patch tokens.
- Optionally concatenate goal image or goal vector embeddings.
- Use a block-causal attention mask: full attention within each frame, causal masking across frames.
- Predict actions from the last patch token of each frame.
Why it matters
- Preserves spatial detail for precise manipulation.
- Keeps the benefits of large-scale visual pretraining.
- Stays compatible with standard transformer policy heads.
The image also shows the method evaluated with state-of-the-art policy heads, and the paper claims strong recovery of specialist scores while keeping inference practical for reactive control.
Related event: Patch Policy Outperforms 7B VLA with 0.7% Parameters(5 posts)→
More from Embodied
- Chums: open-source Apple Watch app puts Claude Code and Codex on your wrist — brandon_galang · 2026-10-27
- Dev uses Grok bot to turn ESP into an iPhone accessory for syncing robot servos — Baconbrix · 2026-10-27
- NVIDIA releases MotionBricks: real-time humanoid motion generation at 15,000 FPS with 2ms latency — _akhaliq · 2026-10-26
- Claude Code finds root backdoor in Honda Civic head unit, then debloats it — IndraVahan · 2026-10-26
- Temple health wearable opens pre-orders, users ask: why wear it before benefits are proven — abhishekcode42 · 2026-10-26
- FLASH: GPU-native deformable sim learns cloth folding in minutes on one RTX 5090 — ChongZzZhang · 2026-10-26