ICLR 2026 · International Conference on Learning Representations

EgoBrainSynergizing Minds and Eyes For Human Action Understanding

Nie Lin1*, Yansen Wang2, Dongqi Han2, Weibang Jiang2*, Jingyuan Li2*, Ryosuke Furuta1, Yoichi Sato1, Dongsheng Li2

1The University of Tokyo2Microsoft Research

* Work done during an internship at Microsoft Research · Corresponding authors

Paper PDF OpenReview Poster
EgoBrain-Mini start here Raw dataset

TL;DR The world’s first egocentric dataset fully synchronized with brain activity: 61 hours of 32-channel EEG and first-person video from 40 people doing 29 everyday actions.

Isometric illustration of the EgoBrain recording room: a seated participant wearing an EEG cap, a head-mounted camera and wrist and arm IMU bands, at a desk with everyday objects. Objects are interactive.
Tap the desk objects, the EEG cap, the camera or the wrist bands
40participants
61hours, synchronized
32EEG channels
29everyday actions

Start here

EgoBrain-Mini in 30 seconds

Pick an action, watch the real recording with its EEG and IMU in sync, then copy the code that loads exactly that segment.

1Pick an action

2See it: video, EEG and IMU on one clock

—

EEG · 4 of 32 channelsIMU · right-wrist acceleration0.0 s

Scalp map

lowhigh

Band power at the 32 electrodes, compared with each other at every moment and interpolated over the head (nose up). It follows the video, four updates a second.

3Load it

Python · this exact segment
Or the whole index
from datasets import load_dataset
ds = load_dataset("ut-vision/EgoBrain-Mini", split="train")
eat = ds.filter(lambda r: r["action"] == "Eat Snack")

The dataset is gated: request access on its Hugging Face page, then run huggingface-cli login once. Only the files of the chosen segment are downloaded.

EgoBrain-Mini on Hugging Face → Browse all 1,562 segments