0.09B DINOv2+MLP beats π0 and OpenVLA on LIBERO, CoRL paper questions benchmarks
anand_bhattad · x · 2026-11-10
- A #CoRL2026 paper by JiangTianchong's team asks: what does a high benchmark score actually tell us about a policy?
- A tiny 0.09B DINOv2+MLP policy with no language model scores 97.6% on LIBERO, beating π0 (94.2%) and OpenVLA (76.5%).
- The paper systematically analyzes the 5 most-reported manipulation benchmarks, questioning their saturation and ability to separate capable policies.
More from Embodied
- ODE Policy learns action dynamics via neural ODEs for multimodal robot trajectories — ntumars · 2026-11-10
- SJTU AutoLab proposes A²-MEAS: correction-code multi-agent framework for embodied AI — AutoLab-SJTU · 2026-11-10
- Robot playing drums? No — the robot *is* the drums — carlosdponx · 2026-11-10
- Perceptual attacks: text tampering tricks every frontier model into harmful robot actions — gabriberton · 2026-11-10
- NVIDIA study: GPT-6 Astra hits 74.8% on RoboLab-120 as agentic robot policy — ZeeshanZiaML · 2026-11-10
- Developer uses a 1.9B decision model on a laptop to flag dirty restaurant tables in real time — JosephJacks_ · 2026-11-10