GitHub repo reproduces Patch Policy for embodied control with DINOv2 features
k7agar · x · 2026-07-28
A GitHub repo reproduces Patch Policy — the paper Efficient Embodied Control via Dense Visual Representations (arXiv:2607.18236).
- Reproduction stack: frozen DINOv2 ViT-S/14 patch features + a Diffusion Policy head.
- Architecture: 224×224 frames are converted into 256 patch tokens (384d), then a block-causal transformer produces per-frame readouts, which feed a DDPM denoiser for a 5-action chunk in [-1, 1].
- Reported results: on Push-T, the repo reaches 0.772 coverage versus the paper target of 0.83 in 100 rollouts.
- The README also shows a SO-101 demo running at 2× speed.
More from Embodied
- Physical AI could make industrial robots easier to deploy — Responsible-Grass452 · 2026-07-28
- Topo, the 1983 programmable home robot that ran code before sensing the world — tobowers · 2026-07-28
- Papers with Code’s robotics hub tracks the field’s main benchmarks and trends — NielsRogge · 2026-07-28
- Vstone unveils a 97 kg mobile dual-arm humanoid robot for factory and logistics work — CyberRobooo · 2026-07-28
- Beijing’s two AI bets: Kimi K3 for digital intelligence and Qianxun for robots — 智东西 · 2026-07-28
- SkoBot turns a toy robot into a language-learning tool for Indigenous kids — danielleboyerr · 2026-07-28