RoboARCade: leaderboard
Agentic robot control: general agents act directly as robot policies through a tool interface on long-horizon simulated manipulation tasks across three embodiments. RoboARCade v1 zero-shot benchmark: 22 long-horizon simulated manipulation tasks (Panda, bimanual Panda and bimanual YAM), five seeds each, averaged over all trials. Success is the share of completed rubric subtasks. Claude models run in Claude Code and GPT models in Codex, all at the highest available reasoning effort (xhigh) with native context compaction. The VIA and multi-view RGBD control interfaces stay separate configurations. Cost, turns and calls per end-effector command are not admitted; the autonomous-improvement arms (Table 2) are separate methods and not admitted. Question access: Task sources are cited public simulation assets; the RoboARCade task release was not verified.
Metric: Rubric subtask success (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Benchmark Sweep Run: source=RoboARCade; runtime=Codex; interface=VIA; effort=xhigh; zero-shot) | 49.8 | |
| 2 | GPT-6 Astra (Benchmark Sweep Run: source=RoboARCade; runtime=Codex; interface=multi-view RGBD; effort=xhigh; zero-shot) | 48.3 | |
| 3 | Claude Fable 5.1 (Benchmark Sweep Run: source=RoboARCade; runtime=Claude Code; interface=VIA; effort=xhigh; zero-shot) | 43.6 | |
| 4 | Claude Fable 5.1 (Benchmark Sweep Run: source=RoboARCade; runtime=Claude Code; interface=multi-view RGBD; effort=xhigh; zero-shot) | 36.8 | |
| 5 | Claude Opus 5 (Benchmark Sweep Run: source=RoboARCade; runtime=Claude Code; interface=VIA; effort=xhigh; zero-shot) | 30.3 | |
| 6 | GPT-5.6 Sol (Benchmark Sweep Run: source=RoboARCade; runtime=Codex; interface=VIA; effort=xhigh; zero-shot) | 14 |
Interactive version: theaggregate.ai/benchmark?slug=roboarcade · How It Works · Data refreshed daily, snapshot 2026-11-07.