26 task types and 794 instances, ranked on a held-out test split.
As of November 17, 2026, Motubrain 2.0 leads with 73.5, followed by PICA-WAM (61.4) and OpenWAM-α (55.3). Official leaderboard ↗ · Paper ↗
Leaderboard
- Tasks
- 26
- Attempts / task
- 20
- Median cost / task
- $0.030
- Median time / task
- 32 s
Held-out test split. Score gives partial credit.
Motubrain 2.073.583.5$0.023est.26 sest.2
PICA-WAM61.474.9$0.025est.28 sest.3
OpenWAM-α55.370.0$0.027est.29 sest.
InternW0-Δ49.266.0$0.028est.30 sest.
GPT-6 Astra46.765.4$17.74est.31 sest.
π0.541.454.4$0.029est.32 sest.10
InternVLA-A1.534.246.8$0.030est.33 sest.11
π033.747.5$0.031est.34 sest.12
GigaBrain-0.733.346.0$0.031est.34 sest.13
Cosmos3-Edge-Policy29.342.5$0.032est.34 sest.14
Fast-WAM25.637.1$0.032est.35 sest.15
X-VLA23.735.8$0.033est.36 sest.Tasks
26 tasks. Hover a task to play its clip (26 of 26).
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ Clip
▶ ClipMethodology
- Scoring
- Code check
- Time cap / attempt
- 40 s
- Control rate
- 60 Hz
- Models in index
- 12 of 15
Scoring. Ordered state conditions in code; score is the share of stages met.
Who runs it. Shanghai AI Lab runs the simulator, the entrant runs the model.
Robots. Simulated arms in Isaac Sim (GenManip).
Note. The operator also makes InternW0-Δ and InternVLA. π0 and π0.5 are EBench fine-tunes.
Time per attempt. Not published. Estimated otherwise: time cap 40 s × (1 − success/2).
Cost per attempt. Not published. Estimated otherwise: GPT-6 Astra's median published cost per attempt on other boards; one H100 at $3.29 per hour (Lambda on-demand, Oct 2026) × time. GPU price source ↗
