Research2026-09-30
Researchers introduced EgoTools, a suite for egocentric tool-use understanding. It includes EgoTools-Data, a 100-hour corpus of tool-centric egocentric videos. It also includes EgoTools-Bench, a 1,000-question diagnostic benchmark. The benchmark covers perception, geometry, procedure, and causal reasoning. Current models struggle with grounding tool use in visual evidence. Fine-tuning on EgoTools-Data improved Qwen3-VL-8B-Instruct from 50.0% to 60.9%. The paper establishes EgoTools as a unified resource for training and evaluation.
What it unlocks
Training and evaluating models on real-world egocentric tool-use understanding.
- 100 hours of egocentric recordings
- 1,000 QA pairs across four tracks
- Gemini-3.1-Pro 66.9% overall accuracy
- Gemini-3.1-Pro 51.7% on Perception & Grounding
Sources