RobotAgentBench Full - Discovery: leaderboard
RobotAgentBench v1 evaluates the AgiBot G1 robot with 30 randomized instances per task/scenario and 50 layouts. Its 11 standalone policies are fine-tuned on a shared 101.8 hours of RAB demonstrations; they are distinct from the released backbones. Agent systems use Codex GPT 6 Astra, Claude Code Opus 5.5, or Antigravity Gemini 4 Argon; hybrids use InternVLA-A1.5. Each episode allows 200 model responses, with 600 control steps per tool call. OpenEnd terminates after 120 simulated seconds without progress. RSI reports the final result after 10 iterations. Scores measure physical process progress, and the task-level score averages five capabilities. The sim-to-real study uses a separate matched subset of 30 instances per setting. Metric: Normalized process score (%).
Metric: Normalized process score (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 18 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | RobotAgentBench Codex + InternVLA-A1.5 C/H/C native robot harness | 37.2 | |
| 2 | RobotAgentBench Antigravity CLI C/C/C native robot harness | 36.5 | |
| 3 | RobotAgentBench Codex C/C/C native robot harness | 35.2 | |
| 4 | RobotAgentBench Codex C/C/C + RSI native robot harness | 33.3 | |
| 5 | RobotAgentBench Claude Code C/C/C native robot harness | 24.5 | |
| 6 | RobotAgentBench Codex + InternVLA-A1.5 –/L/C native robot harness | 4.3 | |
| 7 | RobotAgentBench Codex + InternVLA-A1.5 C/L/– native robot harness | 4 | |
| 8 | RobotAgentBench MolmoAct2 VLA RAB fine-tuned | 1.7 | |
| 9 | RobotAgentBench Galaxea G0.5 VLA RAB fine-tuned | 1.5 | |
| 10 | RobotAgentBench LingBot-VA WAM RAB fine-tuned | 1.5 |
Interactive version: theaggregate.ai/benchmark?slug=robotagentbench-full-discovery · How It Works · Data refreshed daily, snapshot 2026-10-28.