Hub Labs open-sourced HubEgo-100K on Hugging Face in November 2026, releasing 10,124 hours of 1080p egocentric video plus aligned IMU from 12,227 contributors under CC BY 4.0. The HubLabs dataset card describes release 1 as the first stage toward 100,000 hours of skilled manual work for robotics and VLA training.
For physical-AI builders, this is a rare commercial-use-friendly corpus shot in real workplaces, not a scripted lab tabletop.
TL;DR
| Metric | Release 1 (per Hub card) |
|---|---|
| Video hours | 10,124 h |
| Episodes | 72,754 |
| Contributors | 12,227 |
| Task descriptions | 62,000+ across 56 types |
| Formats | LeRobot v3, MCAP, annotations |
| License | CC BY 4.0 |
What Hub recorded
Hub highlights no staged sessions: contributors film their own trades on building sites, in workshops, salons, commercial kitchens, shops, warehouses, and homes.
Hands stay in frame 95%+ of the time, with 96.9% of episode time on active task work, per the card. 34% of hours are labeled hard difficulty and 47% medium.
Sensors include 100 Hz IMU aligned to each video frame plus native streams (accelerometer, gyroscope, magnetometer, device motion, heading, barometer). That pairing matters for teams training policies that fuse vision with motion cues, similar in spirit to miHoYo gamer VLA research but grounded in trades footage.
Privacy processing: Hub says it removed personal and sensitive content before release, and contributors consented to AI and robotics training use.
How to try HubEgo-100K this week
- Open HubLabs/HubEgo-100K and complete the access form.
- Pull episodes in LeRobot v3 if you already train with LeRobot-style stacks.
- Use MCAP paths if your pipeline ingests robotics logs in Foxglove-compatible containers.
- Read LICENSE.md for attribution requirements before shipping commercial models.
- Compare task diversity against Natura-1 dexterous hands coverage when choosing pretraining mixes.
Why open skilled-labor video matters now
Robot gripper mirroring human hand motion for imitation learning from egocentric video
Robot foundation models face a data wall: internet video lacks contact-rich manipulation at scale, and lab demos are cheap but narrow. Hub's pitch, also visible on hub.xyz, is a gig-recorded pipeline with provenance and QA, then open releases for research.
This release is not the full 100k yet. Treat 10,124 hours as stage one while Hub continues contributor payouts.
Crowdsourced egocentric data raises labor ethics questions explainx.ai has noted in robotics supply-chain coverage. The card's CC BY 4.0 terms and contributor consent language are the primary official answers; independent audits are not part of this release.
Episode anatomy and annotation fields
Each episode ships with task description, step-level instructions, action family, skill group, difficulty, and environment labels. That structure helps filter pretraining mixes: you can train bricklaying without salon footage, or oversample hard hours when finetuning contact-rich policies.
Median episode length is about 4.9 minutes with a 30 minute cap per Hub's table. 72,754 episodes across 15,548 environments means many single-environment slices useful for domain adaptation without downloading the full corpus.
IMU alignment at 100 Hz supports models that predict hand motion during occluded moments. Vision-only baselines still benefit, but fusion researchers get a rare open combo at scale.
Training pipeline sketch
Embodied AI robot hand reaching toward objects, echoing contact-rich skilled-labor footage
A minimal LeRobot v3 workflow:
- Request dataset access on Hugging Face and accept CC BY 4.0 attribution terms.
- Install the LeRobot tooling version Hub cites for v3 shards.
- Start with one task type (for example wiring) to validate loaders before multi-task mixing.
- Log train/val splits by contributor id to reduce leakage from the same worker across splits.
- Compare zero-shot grasp metrics against your lab baseline before scaling GPUs.
If you use MCAP, pipe through Foxglove or internal converters to synchronize RGB and IMU timestamps; misalignment silently hurts action labels.
Ethics and labor provenance
Hub states contributors opt in and footage is reviewed. Gig platforms in robotics data always raise wage, consent, and geographic representation questions. The card gives counts (12,227 contributors) but not pay rates in the snippet we cite; treat compensation claims from third-party blogs as unverified unless Hub publishes them on hub.xyz.
For privacy, Hub claims redaction of sensitive content. Deployers remain responsible for bias if certain trades or regions dominate hours.
Builder comparison table
| Dataset angle | HubEgo-100K release 1 | Typical lab demo sets |
|---|---|---|
| Capture | Contributor workplaces | Fixed lab benches |
| License | CC BY 4.0 commercial allowed | Often research-only |
| IMU | 100 Hz aligned | Sometimes video-only |
| Scale today | 10k+ hours open | Smaller public slices |
Task-type examples from the card
Hub lists trades from bricklaying and wiring to phone board microsoldering and watch repair. 56 task types across 30 categories give VLA teams negative examples too: not every hour is kitchen stirring; 83% of hours are workplace settings.
35 action families help you build curriculum learning schedules: start with coarse reach-and-grasp clips before fine motor repair sequences.
Storage and compute planning
1,093,194,072 frames at 1080p is not a laptop download. Plan cloud buckets, partial episode pulls, and WebDataset-style sharding. Hub ships catalog metadata so you can filter before streaming video.
For fine-tuning, pair Hub hours with simulation from robotics scaling discussions rather than expecting hours alone to generalize across grippers.
Commercial use and attribution
CC BY 4.0 allows commercial training if you credit Hub per LICENSE.md. Product teams should legal-review contributor consent language on the card before shipping models trained on egocentric humans.
Frequently asked builder questions
Can I merge HubEgo with my own teleop logs? Yes if licenses align; keep provenance separate in model cards.
Is IMU required? Vision-only baselines can ignore IMU, but you forfeit alignment work Hub already did.
Will release 2 hit 100k hours? Hub markets a staged plan; only 10,124 h are documented today.
Hub versus closed commercial corpora
Closed vendors sell teleoperation and egocentric bundles under NDA. Hub's open slice lets academic labs reproduce baselines without enterprise contracts, similar to how ImageNet unlocked vision. The tradeoff is QA variance across contributor skill.
Compare to closed humanoid demos from Figure-class coverage: demos are polished but narrow; Hub is messy but broad.
Integration with policy and safety work
Physical datasets do not eliminate safety obligations. IEEE-style disclosures on robotics zero-days remind deployers that data scale plus autonomy raises attack surface. Train policies with estop and human override assumptions even when video shows smooth motion.
Contributor geography and diversity
Hub claims 12,227 unique contributors across real workplaces. Check geographic balance in your splits before claiming global generalization. Under-represented trades may need oversampling or synthetic aug carefully labeled.
Model-card language for Hub-trained policies
When publishing weights trained on HubEgo, disclose:
- HubEgo-100K release 1 hours used.
- CC BY 4.0 attribution text.
- Filtering criteria (task types, difficulty).
- Known failure modes (motion blur, lighting, PPE occlusion).
Transparent cards help regulators and customers auditing physical AI products.
Hardware and realism notes
Head-mounted ultra-wide lenses distort depth cues. Policies trained only on Hub may overfit to fisheye motion. Consider calibration steps or crop strategies documented in your eval suite.
PPE, glare, and motion blur appear in real trades footage. Do not assume clean lab lighting transfer.
Pairing Hub data with simulators
Many teams will pretrain on Hub and finetune in Isaac, MuJoCo, or vendor sims. Track sim-to-real gap metrics on contact tasks separately from navigation tasks; Hub emphasizes hands, not locomotion.
LeRobot and Hugging Face access workflow
The dataset card states access is free after a short form. Expect gated downloads typical of large HF datasets. Plan automation with HF tokens and revision pins so nightly training jobs do not break when Hub uploads release 2.
Document which HubLabs versus Hubdata mirror you used; both cards describe the same release at the time of writing, but teams should pin one canonical URL in reproducibility logs.
When mixing Hub video with proprietary teleop, keep license boundaries explicit in composite datasets you redistribute.
Downstream benchmarks to watch
After fine-tuning, evaluate on contact-rich suites separate from navigation leaderboards. Hub's hard and medium tags help stratify evals so you do not claim wins on easy kitchen clips alone.
Share failure clips publicly when safe; robotics benefits from negative results on egocentric blur and occlusion.
If your lab lacks humanoid hardware, start by training perception and intent models on Hub, then distill into policies you later fine-tune on your own gripper stack.
Track hours watched during training, not only loss curves, so you can report data scaling experiments credibly in papers and model cards. Reviewers increasingly ask for data filters, not only parameter counts.
Citation snippet for papers
Use the dataset bibtex Hub publishes:
@dataset{hubego100k2026,
author = {{Hub Labs (Hub Data, Inc.)}},
title = {HubEgo-100K},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/HubLabs/HubEgo-100K}
}
Related reading
- Natura-1 dexterous hands from human video
- DYNA-2 world-action robotics scaling
- Envisioning agentic robotics survey
- Vicone zero-day robotics vulnerabilities
- Embodied AI exposure benchmark
- Unitree and humanoid deployment context
- HIW-500 humanoid teleop dataset context
Sources: HubLabs/HubEgo-100K, Hub Labs site.
