Eidon AI

Eidon built a wearable upper-body rig for capturing human manipulation: a head-mounted camera plus a seven-point IMU harness. We deployed it across ordinary household work like laundry, cleaning, dishes and cooking and countries like Phillipines, Nigeria, and Vietnam.

The company has wound down. Everything we collected is published here under CC-BY-4.0, as a contribution to the community.

How the release is organised

The data is split across three artifacts / datasets. They are different parts of the same data corpus.

	What's inside	Size	Load with
🎥 tracker-pov	The tracker videos. 13,451 egocentric MP4s + one metadata row each	9.05 TB	`load_dataset`
📈 tracker-pov-imu	The sensor data files. 779M rows of 24 Hz orientation + motion for the same recordings	9.5 GB	`load_dataset`
📼 egocentric-pov	Standalone egocentric video with no sensor data	~1.75 TB	`hf buckets sync`

The first two are one dataset in two pieces

tracker-pov and tracker-pov-imu cover the same 13,451 recordings. They join on recording_id. Use both.

from datasets import load_dataset

video = load_dataset("eidon-ai/tracker-pov", split="train")                  # 9.05 TB
imu   = load_dataset("eidon-ai/tracker-pov-imu", split="train", streaming=True)  # 9.5 GB

Filter the metadata first. It is small, and it carries the QC scores, task labels and sensor coverage flags you will want to select on:

import pandas as pd
meta = pd.read_parquet("hf://datasets/eidon-ai/tracker-pov/recordings/metadata.parquet")
good  = meta[(meta.qc_status == "valid") & (meta.has_chest)]

What's in it

1,274 hours of paired egocentric video and seven-point arm tracking, from 27 contributors in their own homes. Every recording passed an automated QC pass (hand presence, lighting, blur, camera stability) and ships with its scores.

Read the limitations section before training on it. The short version: the task distribution is heavily skewed toward folding laundry, contributors are few and unevenly weighted, and you should split by contributor_id, not randomly.