New robot policy runs on an RTX 5090, beating OpenVLA-OFT while using 0.7% of its size
Jeff Cui shows that a frozen ViT backbone plus a tiny transformer can outperform a 7B‑parameter VLM in robot control.
By Ryan Merket · Published
Primary source: X
Why it matters
Patch Policy proves that dense visual features from a frozen ViT can replace massive VLMs for robot control, dramatically cutting compute costs and latency while improving performance.
A lean visual policy beats a 7B‑parameter VLM
Jeff Cui (@jeffacce) announced on July 22nd, 2026 a new robot‑control method he calls Patch Policy. The approach freezes a state‑of‑the‑art Vision Transformer (ViT) pretrained on large‑scale image data, feeds every patch token to a lightweight transformer policy, and adds a simple block‑causal mask. The result, Cui writes, “beats OpenVLA‑OFT with 0.7% of its params, and trains on a 5090.”…
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.