HELM Robo-Reward-Bench - Robo Arena: leaderboard
The "Robo Arena" subtask of HELM Robo-Reward-Bench. Stanford HELM's Robo-Reward-Bench: vision-language models act as reward models over robot manipulation episodes, predicting task progress or success from video.
Metric: Absolute Error. Source: crfm.stanford.edu. Saturation forecast: Around July 2027. 25 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (2026-04-23) | 0.77 | #24 |
| 2 | GPT-5.5 (xHigh) | 0.78 | #29 (GPT-5.5) |
| 3 | GPT-5 Mini (2025-08-07) | 0.86 | #174 |
| 4 | Gemini 2.5 Pro | 0.94 | #153 |
| 5 | GPT-5.2 | 0.97 | #113 |
| 6 | GPT-5 | 1.03 | #104 |
| 7 | Gemini 3 Flash (Preview) | 1.06 | #84 |
| 8 | GPT-5.1 (2025-11-13) | 1.07 | #105 |
| 9 | Gemini 2.5 Flash Lite | 1.14 | #421 |
| 10 | Gemini 2.5 Flash | 1.19 | #242 |
| 11 | Gemini 3 Pro (Preview) | 1.23 | #66 |
| 12 | GPT-5 Nano | 1.29 | #428 |
| 13 | Llama 4 Maverick Instruct | 1.55 | #460 |
| 14 | Llama 4 Scout Instruct | 1.83 | #542 |
No result here: #4 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=helm-robo-reward-bench-robo-arena · How It Works · Data refreshed daily, snapshot 2026-10-23.