RobotEQ-Video - Real Videos: leaderboard

Metric: Real-subset Avg (%; action-properness judgment: given an egocentric robot-view video and a numbered list of candidate robot actions, the model labels each action proper or improper for socially appropriate proactive assistance; score is the mean of accuracy and macro-F1 against the majority-voted human label and of probability of agreement and consensus-weighted agreement with the four individual annotators; 818 real egocentric clips, 5,625 actions). Source: arxiv.org. Saturation forecast: Around March 2028. 48 models tracked.

Top models

#ModelScore
1GPT-5.6 Luna79.9
2GPT-5.578.7
3Gemini 3.5 Flash78.1
4Seed 2.0 Lite78.1
5Qwen 3.5 27B (Thinking)77.5
6Gemini 3.1 Flash Lite77
7Kimi K2.675.5
8GLM-4.1V-9B (Thinking)75.3
9Gemma 4 12B (IT)75.3
10Qwen 3 VL 30B A3B Instruct75.2
11Qwen 3 VL 8B (Thinking)74.6
12Gemma 4 31B (IT)74.5
13Qwen 3 VL 8B Instruct74.5
14Qwen 3 VL 4B (Thinking)74.5
15Qwen 3 VL 30B A3B (Thinking)74.2

Interactive version: theaggregate.ai/benchmark?slug=roboteq-video-real-videos · How It Works · Data refreshed daily, snapshot 2026-09-26.