DexBench: leaderboard
A robot policy, fine-tuned on DexBench's demonstrations, drives a two-armed robot with two five-finger hands through 15 simulated industrial tasks (retail, logistics, manufacturing, hotel service); the number is the site's Mean SR, the unweighted mean of the task success rates over 46 cells of 60 frozen start states (2,760 episodes per policy). Unlike LIBERO's tabletop gripper tasks it needs dexterous hands and scores held-out objects and placements beyond the training range.
Metric: Mean success rate over 15 dexterous manipulation tasks, 60 frozen episodes per cell, v1.0 (%). Source: dexbench.ai. Saturation forecast: Rough model projection: around 2026. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Fast-WAM-Joint | 20.7 | |
| 2 | RLDX-1.1-PT-IMG | 16.5 | |
| 3 | Cosmos3-Nano | 16.2 | #295 |
| 4 | GR00T-N1.7 | 8.6 |
Interactive version: theaggregate.ai/benchmark?slug=dexbench · How It Works · Data refreshed daily, snapshot 2026-11-13.