VA-Bench - Single-Arm: leaderboard

Metric: Task success (%; the 11 single-arm base task families; 20 physically verified seeds per task family in RoboTwin simulation; the model watches an RGB-only expert video, writes its own task summary, chooses camera viewpoints and issues metric Cartesian arm commands that a fixed model-agnostic controller executes, with execution feedback; terminal success from the environment checker; task-family macro-average, mean of three runs). Source: arxiv.org. Saturation forecast: Around December 2026. 12 models tracked.

Top models

#ModelScore
1Qwen 3.8 Max65.61
2Claude Opus 564.09
3GPT-5.6 Sol59.55
4GPT-5.6 Terra29.24
5Seed 2.1 Turbo19.09
6Gemini 3.6 Flash18.03
7GPT-5.6 Luna17.42
8Claude Sonnet 511.52
9Qwen 3.7 Max10.3
10Qwen 3.7 Plus9.09
11MiniMax-M35.15
12MiMo-V2.54.24

Interactive version: theaggregate.ai/benchmark?slug=va-bench-single-arm · How It Works · Data refreshed daily, snapshot 2026-09-26.