Patch Policy beats OpenVLA-OFT with 0.7% of the parameters on one RTX 5090
ChongZzZhang · x · 2026-07-23
Patch Policy argues that a robot policy does not need 7B parameters if it can use dense visual features.
- It combines a pretrained ViT with a small transformer and reportedly beats OpenVLA-OFT while using only 0.7% of the parameters.
- The policy was trained on a single RTX 5090.
- In the demo, the system inserts a cable with about 2 mm tolerance and succeeds again even when the rollout is interrupted mid-execution.
More from Embodied
- Tesla Robotaxi Logs Over 380,000 Miles With Zero Notable Incidents — yunta_tsai · 2026-07-23
- Lightwheel launches SimReadyGen for physics-grounded 3D simulation assets — Sentdex · 2026-07-23
- Tesla plans Terafab fab to speed Optimus AI chip development — XFreeze · 2026-07-23
- NVIDIA says physical AI needs exabyte-scale simulation data at SIGGRAPH 2026 — jonstephens85 · 2026-07-23
- Musk says Optimus aims to be the first humanoid robot useful in daily life — XFreeze · 2026-07-23
- NVIDIA says its open-source robotics simulator cut training from 5 hours to under 2 minutes — imjustnewatai · 2026-07-23