ATOM-Bench - Single-Arm Composition: leaderboard
Held-out compositional manipulation on a real Franka Panda arm from ATOM-Bench (Wu et al., 2026): 12 tasks that each combine motor skills (pick-and-place, reorienting, pushing, stacking, pouring, opening a drawer) with two or more language constraints (color, shape, size, count, exclusion, spatial relation, destination), none of them demonstrated during fine-tuning. The policy is fine-tuned only on the benchmark's atomic tasks; success rate in percent over 10 fixed physical test seeds per task, macro-averaged over tasks; higher is better.
Metric: Success rate (%; held-out Composition Set, 12 tasks x 10 fixed physical test seeds on the Franka Panda single-arm track, macro-averaged over tasks; one policy fine-tuned on the 15 atomic tasks of the track). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Pi0.5 | 15.8 | |
| 2 | Motus | 10.8 | |
| 3 | LingBot-VLA | 3.3 | |
| 4 | GR00T-N1.6 | 3.3 | |
| 5 | SmolVLA | 0 |
Interactive version: theaggregate.ai/benchmark?slug=atom-bench-single-arm-composition · How It Works · Data refreshed daily, snapshot 2026-11-12.