Eidon Tracker POV
1,274 hours of egocentric household video with seven-point arm IMU tracking
License
CC-BY-4.0
Overview
Twenty-seven volunteers wore a head-mounted camera and a seven-sensor IMU harness while doing real chores in their own homes — folding laundry, cleaning, dishes, cooking, knitting. The result is 13,451 recordings totalling 1,273.8 hours, each pairing first-person video with 24 Hz orientation for both hands, both forearms, both upper arms and the chest. The chest sensor is the reference frame: composing its inverse with a limb gives torso-relative arm pose, invariant to which way the wearer is facing. Video and IMU live in two repositories joined on recording_id, with a third bucket holding a further 306 hours of video-only footage. All of it is CC-BY-4.0.
Large robot-learning datasets such as AgiBot World, DROID, Open X-Embodiment and ABC-130k record robots. This one records people doing ordinary manipulation, from their own point of view, with their arm configuration measured rather than inferred. Estimating hand pose from egocentric video is a hard, noisy problem; here the arm poses come with ground truth from body-worn sensors.
The dataset card documents its own weaknesses in detail. The task distribution is heavily skewed: folding laundry is 67.5% of all hours and the tail classes have under three hours each, so the card warns against treating it as a balanced nine-class activity dataset. Contributors are few and unevenly weighted, the largest supplying 16.1% of hours and the top five 57.2%, so the card advises splitting on contributor_id: a random split leaks the same person, home and camera rig across train and test and inflates results. Demographics are disclosed and described as unrepresentative: of those who reported, the median age is 26 and 15 of 17 are male. Raw accelerometer, gyroscope and magnetometer readings cover only 2,841 recordings, and the card names the exact field to filter on. Quality control ran MediaPipe hand detection, HSV brightness, Laplacian blur and optical-flow stability checks on every recording; the 472 rejected recordings are included and marked, as negative examples.
The company behind it, Eidon AI (trading as Solidic Labs Inc), has wound down. It released the data under CC-BY-4.0 so that the data would outlive the company, which makes this a final release rather than a first one with more promised.
Why it matters
Most large robot-learning datasets record robots being teleoperated. Policies that learn from human video need human video, and the hard part has always been knowing what the arms were actually doing: pose estimation from a first-person camera is unreliable at the wrist and worse at the shoulder. Recording 1,274 hours of real household work with seven IMUs on the body answers that directly, and the data is released under a permissive licence.
Problem it solves
Learning manipulation from human video needs to know what the arms were doing, and recovering that from a first-person camera is unreliable. This records it directly: 24 Hz orientation from seven body-worn sensors alongside the video, over a thousand hours of unstaged household work, free to use commercially.
Key takeaways
- 13,451 recordings, 1,273.8 hours, 27 contributors, 9.05 TB of MP4 at 1080p to 4K
- Seven 9-axis IMUs at 24 Hz — both hands, both forearms, both upper arms and the chest, with the chest acting as a torso-relative reference frame
- Video and IMU are separate repositories joined on recording_id; a third storage bucket adds 306 hours of video-only footage from 37 contributors
- Real chores in real homes rather than staged demonstrations: laundry, cleaning, dishes, cooking, drawing, knitting, bed-making, plant-watering
- Every recording carries QC metrics and a valid/flagged/invalid status; the 472 invalid ones are included deliberately as negative examples
- The dataset card names the pitfalls: skewed tasks, contributor leakage on a random split, sparse demographics, and which metadata field actually indicates raw motion data
- CC-BY-4.0, commercial use permitted with attribution, published as a final release after the company behind it wound down
Who it is for
- Anyone training manipulation policies or world models on human demonstration video
- Researchers working on egocentric hand and arm pose estimation who need ground truth to evaluate against
- Teams building activity recognition for activities of daily living, in assistive robotics or elsewhere
- People studying sensor fusion between vision and body-worn IMUs
- Anyone who wants scale in household manipulation data without collecting it themselves
Limitations
- Folding laundry is 67.5% of all hours; the tail activities have under three hours each and cannot carry a class on their own
- Twenty-seven contributors with the top five supplying 57.2% of hours — split on contributor_id or your evaluation will leak the same person, home and rig into both sides
- Demographics are sparse and skewed: of those reporting, median age 26 and 15 of 17 male
- Raw accelerometer, gyroscope and magnetometer data covers only 2,841 of 13,451 recordings; orientation quaternions are present throughout but the raw streams are opt-in
- The video is unprocessed — no face blurring or redaction, recorded inside private homes, and the card is explicit that incidental capture of faces, documents or screens cannot be ruled out
- 9.05 TB for the video alone, so this is not a dataset you casually pull down
- No hardware synchronisation between camera and IMU; alignment is by timestamp, with an offset the authors describe as a few milliseconds of sequential code execution
- The organisation that produced it has wound down, so there will be no corrections, no v2 and nobody to answer questions beyond the removal-request address on the card