RoboWorld - Test Reward: leaderboard
RoboWorld (arXiv 2610.40565): coding agents in Claude Code or Codex control simulated arms, a mobile manipulator and humanoids through code, three hours per attempt, scored by a verifier on the physical simulation state. Average partial-progress reward (0-1) on the 25 test tasks, where every model so far has 0% success.
Metric: Average reward (0-1) on the 25 test tasks. Source: arxiv.org. Saturation forecast: Around 2031. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Opus 5.5 (xHigh) | 0.06 | #3 (Claude Opus 5.5) |
| 2 | GPT-6 Astra (xHigh) | 0.01 | #6 (GPT-6 Astra) |
| 3 | GPT-6.1 Terra (xHigh) | 0 | #15 (GPT-6.1 Terra) |
No result here: #7 GPT-6.1 Sol, #9 Gemini 4 Argon, #10 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=roboworld-test-reward · How It Works · Data refreshed daily, snapshot 2026-11-05.