LeRobot — Eidon AI winds down and open-sources 9TB egocentric housework dataset (tracker-pov) with 7-point IMU arm tracking

TL;DR

LeRobot announced on Sep 21, 2026 that @eidon_ai — one of the 2026 first-person data-capture startups — is winding down and has open-sourced its full corpus on Hugging Face at huggingface.co/datasets/eidon-ai/tracker-pov under CC-BY-4.0. The release is 9TB total: 1,274 hours of first-person housework video across 13,451 recordings from 27 homes, paired with 7-point IMU-based arm tracking. Community comment (@AIWithPrince) flags that 67% of the corpus is laundry folding, positioning it as an unusually deep benchmark for deformable-object manipulation. A distinctive datapoint on the release-shape axis: not a fresh launch but a company-shutdown data drop — the first filed instance of this pattern in the egocentric-data cluster.

Key claims

Eidon AI is winding down as a company and open-sourcing its accumulated data under CC-BY-4.0, published on Hugging Face at eidon-ai/tracker-pov [tweet body].
Corpus scale: 1,274 hours of first-person video, 13,451 recordings, 27 homes, 9TB total [tweet body].
Modality shape is video + 7-point IMU arm tracking (not full body pose, not hand pose) — a distinctive combination vs. camera-only egocentric releases and vs. full-body-pose annotated releases [tweet body].
Community observation: ~67% of the corpus is laundry-folding footage, making it an unusually deep specialized dataset for deformable-object manipulation rather than a broad-domain corpus [@AIWithPrince reply].

Method

Not a research paper — this is a company-shutdown data release announcement. From the tweet + linked HF dataset URL:

Capture: first-person / egocentric video collected in real homes across 27 households.
Sensing: paired 7-point IMU stream for arm tracking (specific joint layout not disclosed in the tweet).
Domain: housework tasks, heavily skewed toward laundry folding per the top community reply.
Distribution: Hugging Face dataset, CC-BY-4.0 (commercial use permitted with attribution).
Framing: released because the company is winding down, not as a launch milestone — the 9TB / 13,451-recording / 27-home scale represents whatever Eidon AI had collected before shutting operations.

Results

No model / no benchmark numbers — this is a raw data release from a wind-down. Signals worth tracking:

9TB scale, 1,274 hours, 27 homes, 13,451 recordings [tweet body].
11.8K tweet views within a few hours of posting [tweet metadata].

Why it’s interesting

Adds a distinct release-shape datapoint to the Open foundation-model releases page: the shutdown data drop, previously not filed. Companies winding down and releasing their data corpus under permissive licenses is a new release primitive alongside the “coordinated flagship launch” (NVIDIA / Kimi / Qwen), the “capture-hardware-as-open-release” (Ego-OSCAR: Egocentric Open source Stereo CAptuRe System), and the “process documentation as release primitive” (Smol Training Playbook (GPU MODE talk on SmolLM3)) already tracked there — worth watching whether the pattern repeats as more first-person-data startups run out of runway. On the data side, sits at a specific point in the current egocentric-release landscape: much smaller than Lightwheel AI open-sources EgoSuite-Open100K — 100K-hour egocentric human dataset with hand and body pose EgoSuite-Open100K (100K h with hand/body pose) or Egocentric-1M — largest egocentric video dataset (Build AI / Eddy Xu announcement) Egocentric-1M (~1M h factory footage), and much smaller than RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data RekaDaily-10k (10K h household). But deeper on a specific task: 67% laundry folding at 1,274 h is likely more concentrated deformable-object footage than any other filed release. The 7-point IMU arm tracking is also a distinct sensor combination — most egocentric releases ship either raw video (Build AI’s Egocentric-1M) or full body pose annotations (Lightwheel’s EgoSuite), not IMU-only wearable-motion capture — which makes it plausible pretraining data for arm-focused VLA/WAM policies where wrist/forearm trajectories matter more than full body kinematics. Complements HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining‘s finding that egocentric human data with instrumentation outperforms raw robot data at fixed budget: Eidon’s IMU signal is precisely the kind of instrumentation that argument called for.