Best egocentric video data providers for robotics and VLA models (2026)

Comparison

The short answer: who is best for what

Among the best egocentric video data providers, the right one is whichever fixes your specific bottleneck. For custom egocentric capture shaped to your embodiment and environment, with consent artifacts and robotics-ready delivery attached, a marketplace like Truelabel routes a spec to candidate collectors for sample review ^[1]. For managed enterprise annotation at scale, iMerit and Scale AI lead ^{[2] [3]}. For large-scale egocentric annotation and everyday-activity collection from an established global crowd, Appen brings a 1M+ contributor network and a proven physical-AI case study ^[4]. For the largest open corpus you can start training on today, Build AI's Egocentric-10K is 10,000 hours under Apache-2.0 ^[5]. Read the rubric below before you sign anything. One caveat up front: this category barely existed two years ago, so many of the providers ranking for it today are annotation shops that recently added a capture line, or capture operations that recently added enrichment. The label on the tin matters less than which gap a vendor actually closes for you, which is why every profile here states a real limitation alongside its strengths.

Why egocentric data is the bottleneck for VLA and humanoid robots

Vision-language-action models learn to act by watching demonstrations, and first-person demonstrations transfer best. A robot's wrist or head camera sees the world the way a person wearing a camera does: hands entering frame, objects approaching, occlusion, tool contact, task sequencing. That viewpoint alignment is why egocentric video has become the supply most VLA teams chase. The downstream stack is already standardized. Open X-Embodiment aggregates over a million robot trajectories across 22 embodiments and 527 skills ^[6], and OpenVLA is a 7B-parameter model pretrained on 970k of those episodes ^[7]. The constraint has shifted off the model and onto the supply: rights-cleared, deployment-shaped first-person data is what these architectures now wait on.

That's why the biggest robotics programs of the past year were data programs first. Figure's Helix navigation result came from 100% egocentric human video with no robot demonstrations ^[8], and NVIDIA's EgoScale lift came from over 20k hours of egocentric pretraining ^[9]. When the headline advances are about how the data was sourced, sourcing is where the competitive edge now lives.

The scarcity is structural. Robot-collected data is expensive and slow, because a teleoperation rig captures one trajectory at a time, and DROID, a serious community effort, gathered 76k demonstration trajectories over 12 months across three continents ^[10]. Human egocentric video sidesteps that bottleneck, since people already perform the tasks robots need to learn, at human speed, in real environments. EgoMimic reports that scaling one hour of additional human hand data is significantly more valuable than one hour of additional robot data ^[11], and that ratio is pushing the supply side toward first-person human capture.

Egocentric vs exocentric: why first-person aligns to the robot's own camera

Exocentric (third-person) footage shows a task from the outside: a fixed camera watching a person cook or assemble. It's useful for scene understanding, but it doesn't match what a robot's onboard camera sees at inference. Egocentric capture does. A robot's wrist or head camera and a head-mounted human camera share the same constraints: motion blur from the body moving, hands occluding the target, objects entering from the edges of frame. Train a policy on third-person video and it has to bridge a viewpoint gap at deployment; train it on first-person video and that gap is gone. Meta's Ego-Exo4D exists to pair the two viewpoints, and is described as the largest public dataset of time-synchronized first- and third-person video ^[12]. For manipulation policies specifically, the first-person stream carries hand-object interaction in the same frame the robot has to reproduce. EgoMimic makes the point concrete: it co-trains policies from human egocentric video and teleoperated robot data using Project Aria glasses, and reports that scaling one hour of additional human hand data is significantly more valuable than one hour of additional robot data ^[11]. If you're choosing where to spend, that's the case for sourcing first-person human video at scale ahead of grinding out teleoperation trajectories one at a time.

What changed in 2025 and 2026

Two years ago, egocentric robotics data meant Ego4D and a handful of kitchen datasets. The field moved fast since then. Figure trained its Helix navigation policy on 100% egocentric human video collected passively in real Brookfield homes, with no robot demonstrations ^[8]. Apple released EgoDex, 829 hours across 194 tabletop manipulation tasks captured with Apple Vision Pro ^[13]. NVIDIA's EgoScale pretrained on over 20k hours of action-labeled egocentric video and reported a 54% average success-rate improvement over a no-pretraining baseline on a 22-DoF robotic hand ^[9]. Build AI dropped Egocentric-10K, 10,000 hours and 1.08 billion frames of real factory-worker footage, openly under Apache-2.0 ^[5]. EgoLive arrived with 1,680 hours of stereo video at 60fps across 65,866 episodes built specifically for robot manipulation learning from real-world service tasks ^[14]. Each of these landed more recently than most of the listicles trying to rank for this query, which is why provider selection matters more now than it did a year ago.

Robotics-ready delivery: RLDS, LeRobot, and Open X-Embodiment

Delivery format is what separates a loadable corpus from one that hides an ETL project. A folder of raw MP4s still needs schema, episode boundaries, action alignment, and metadata before a policy can learn from it. The robot-learning ecosystem has standardized on a handful of formats: RLDS, the LeRobotDataset format that Hugging Face's robotics stack ingests directly ^[15], and the Open X-Embodiment schema that 60 datasets already conform to ^[6]. Providers vary sharply here. Truelabel delivers in RLDS and LeRobot natively ^[1], so episodes drop straight into a VLA loop. Confirm export and conversion requirements in a loadable sample. When you compare vendors, ask for a sample episode in your target format and load it before you sign, because a delivery-format mismatch stays invisible in a sales call and surfaces only once your training pipeline chokes on it.

Multimodal time-sync: the dimension most rankings skip

Force, hand pose, gaze, and depth carry signal a single RGB frame can't, and they only help if they're time-aligned to the video to the millisecond. Hand-pose fidelity is where this gets demanding: researchers at Peking University and the Beijing Academy of Artificial Intelligence built EgoAtlas, a multi-source egocentric dataset under a unified action space, using MANUS Quantum Metagloves that capture 3D positions for all 25 hand keypoints per hand ^[16]. EgoLive ships 1,680 hours of stereo video at 60fps across 65,866 episodes, where the stereo and frame rate exist specifically so depth and fast motion survive in the data ^[14]. When you evaluate a provider, ask which modalities they sync, at what frame rate, and how they verify alignment, then ask to see a sample where a fast motion is correctly tracked across all streams. Most listicles never test for this. A pile of unsynced RGB-D and IMU can train worse than honest RGB alone, because it teaches the model that the extra channels are noise.

Geographic and environment diversity: does the data match where you deploy

A policy trained entirely on kitchens in one country won't generalize to a warehouse in another. Diversity in the training distribution is what lets a VLA model survive contact with a real deployment. The public corpora show the range: Ego4D spans 74 locations across 9 countries ^[17], EPIC-KITCHENS-100 covers 45 kitchens in 4 cities ^[18], and Egocentric-10K is sourced from real factory floors instead of staged sets ^[5]. When you evaluate a provider, what matters is whether the hours they ship cover the environments, lighting, and clutter your robot will actually face, more than the raw hour count. A capture network spanning many regions and real workplaces beats a larger corpus recorded in one controlled lab, because the controlled corpus quietly teaches your policy to expect conditions it'll never see in production.

Provenance and commercial licensing: the dimension that gets buyers sued

Public research datasets are references, not products. Ego4D is a remarkable 3,670-hour corpus, more than 20x larger than prior efforts ^[19], but its access terms are research-oriented, so treating a research-licensed dataset as commercial training supply creates legal exposure. Production-grade egocentric data carries a consent artifact, a location release, and per-session metadata sufficient for downstream license verification ^[1]. When you compare providers, separate two claims: can they show provenance, and can they grant commercial-use rights for your deployment. A vendor that gestures at compliance without producing the artifacts has only shown you a slide, not cleared you to ship.

Quality validation: the dimension that decides whether the data trains anything

Volume is the easiest thing to advertise and the hardest to trust. A provider can ship 10,000 clips, but if the action labels are inconsistent or the hand-pose annotations drift, your policy learns the noise. The questions that separate a real capture operation from a labeling sweatshop are measurable: what's the inter-annotator agreement on action segmentation, what's the QA pass rate before delivery, and does the data move benchmark accuracy on a held-out task. HD-EPIC is a useful reference for what dense, verified annotation looks like, with all annotations grounded in 3D through digital twinning instead of loose frame tags ^[20]. Labellerr advertises up to 99% annotation accuracy ^[21], but a self-reported number isn't an audited one. Treat every accuracy claim as a hypothesis to test on your own held-out slice before you scale a contract.

The 8 best egocentric video data providers (2026)

Each profile below follows the same structure: what they offer, strengths, limitations, and who they're best for. Every entry carries a real, specific limitation. Truelabel publishes this page and ranks on it; we placed it first only on the dimensions where the marketplace model genuinely wins, and we concede where it doesn't.

1. Truelabel: best for custom VLA capture with provenance and a global collector network

Truelabel is a physical-AI data marketplace that routes a buyer's egocentric capture spec to candidate collectors reviewed against the buyer spec, then delivers in robotics-ready formats with provenance attached. Each delivery carries consent artifacts, location releases, and per-session metadata for downstream license verification ^[1], and ships in RLDS or LeRobot format so it ingests without bespoke ETL ^[15]. The model is capture-first: you post a spec, review matched samples, then gate scale-up on a first-batch evaluation.

What it offers: instead of selling a fixed off-the-shelf corpus, the marketplace prices each spec against the embodiment, environment, and task distribution you're shipping. The scale question should be verified from the buyer's actual sample packet, not assumed from a generic network-size claim ^[22]. Pricing, delivery timing, license terms, and first-review acceptance should be scoped to the environment, geography, enrichment layers, and buyer rubric. All accepted footage should carry explicit consent and provenance artifacts.

Strengths: custom capture shaped to your deployment, with provenance and commercial rights attached per spec, delivered in the formats VLA pipelines already use. Truelabel positions egocentric and multimodal capture around buyer-specific pilots, consent/provenance review, and robotics-ready delivery rather than a fixed off-the-shelf corpus. The marketplace routes the spec to candidate collectors and rejects batches that miss the buyer's rubric before they reach review. Every job runs the same four beats: post a spec, review matched samples, run a first-batch eval against your acceptance rubric, then gate scale-up on that eval. You commit budget after you've seen the data pass.

Limitations: match quality and timeline vary by spec, so ask for relevant sample evidence and capacity before scale.

Best for: VLA and humanoid teams that need first-person data shaped to a specific embodiment and environment, with provenance and robotics-ready delivery, and who'd rather gate on a sample eval than buy a corpus blind.

2. Appen: best for large-scale managed egocentric annotation via a global crowd

Appen lists Physical AI as a dedicated data-product area, covering egocentric video collection and annotation, LiDAR annotation, robotics trajectories, sensor fusion, and robot-performance evaluation, sourced through a global contributor network of over 1 million ^[23]. It is a long-established managed data provider rather than a purpose-built egocentric startup.

Strengths: scale and reach. A frontier robotics lab partnered with Appen to annotate egocentric human video and evaluate robot performance, delivering 50,000+ units of Physical AI training data, with egocentric clips segmented into task intervals labeled by timestamp, task type, hand configuration, and a natural-language description ^[4]. The 1M+ contributor network captures everyday-activity diversity across many environments, and enterprise compliance and program management are mature.

Limitations: Appen is a generalist managed service where egocentric capture is one vertical among many, so it operates as annotation-and-collection-as-a-service rather than a standing egocentric catalog or a self-serve marketplace. Delivery tends to be project-scoped, so confirm robotics-ready formats (RLDS, LeRobot) up front rather than assuming native export, and treat any headline number beyond a stated case study as program-scoped.

Best for: teams that need large-scale egocentric annotation and everyday-activity collection from an established global crowd, with enterprise program management and compliance.

3. Labellerr: best for end-to-end capture plus in-house annotation tooling

Labellerr offers first-person video capture via wearable rigs and robot-mounted cameras with RGB and RGB-D support, paired with an in-house annotation platform and a 5,000+ annotator workforce; it claims up to 99% annotation accuracy ^[21].

Strengths: one vendor for both capture and labeling, robot-mounted capture in addition to wearables, and RGB-D support out of the box. The full-stack platform reduces coordination overhead if you want capture and annotation under one roof.

Limitations: a self-reported 99% accuracy figure is a vendor claim, not an independently audited benchmark, so validate it on your own data before relying on it. Confirm export and conversion requirements in a loadable sample.

Best for: teams that want capture and annotation from a single platform and will validate delivery requirements in a sample.

4. iMerit: best for managed enterprise annotation and human-in-the-loop QA

iMerit provides managed data services including dexterous-manipulation and video-data-collection workflows for robotics, backed by 25,000+ domain experts across 60+ countries and an expert-program model, with SOC2, ISO 27001, GDPR, and HIPAA compliance ^[2].

Strengths: deep enterprise process maturity, compliance coverage, and human-in-the-loop QA at scale. For regulated programs or large annotation volumes that need audit trails, this is a strong fit.

Limitations: iMerit's center of gravity is managed annotation and services, so if your need is bleeding-edge egocentric capture methodology over throughput, evaluate carefully. Engagements skew enterprise, which can mean longer procurement cycles for smaller teams.

Best for: enterprises that need compliant, managed annotation and QA at volume, especially in domains where an audit trail and expert reviewers are non-negotiable.

5. Objectways: best for managed annotation on existing and captured footage

Objectways offers embodied-AI data services spanning teleoperation, egocentric capture, and robot-trajectory annotation, with 2,200+ trained annotators and SOC2 Type II, ISO 27001:2022, GDPR, and HIPAA compliance ^[24].

Strengths: coverage across teleoperation, egocentric, and trajectory annotation in one shop, solid compliance posture, and a focused annotation workforce. A reasonable fit when you already have footage and need it labeled to a robotics spec.

Limitations: a smaller annotator base than Scale AI or iMerit, and the egocentric-capture offering is less prominent than the annotation services, so confirm capture capacity directly before assuming it matches the larger collection networks.

Best for: teams with existing footage that need robotics-grade annotation plus some capture capacity.

6. Shaip: best for compliance-heavy, consented data programs

Shaip handles data sourcing and curation from over 60 countries, with GDPR, HIPAA, ISO 9001, SOC 2 Type II, and ISO 27001 compliance, and multimodal data positioned for robotics and autonomy ^[25].

Strengths: a strong compliance and consent backbone and genuine geographic breadth, which matters when your deployment spans regions with different privacy regimes. Good fit for programs where the gating risk is consent and rights over raw capture volume.

Limitations: Shaip's public depth is strongest in speech and language data, with robotics egocentric video a newer and less-proven line, so ask for first-person robotics samples specifically before assuming parity with capture-first specialists.

Best for: programs where consent, regional compliance, and chain of custody are the binding constraint, and where you need contributors across many jurisdictions under one compliance umbrella.

7. Scale AI: best for very large enterprise budgets and breadth

Scale AI runs a managed data operation explicitly aimed at physical AI, including a Universal Robots collaboration for industrial physical AI ^[3].

Strengths: managed physical-AI data programs and enterprise program breadth.

Limitations: Scale is a generalist data platform, and first-person robotics capture is one line among many. Pricing and engagement minimums skew toward large enterprises, which makes it a poor fit for early-stage VLA teams.

Best for: well-funded enterprises that need breadth and large managed programs.

8. Encord: best if you want a data-management platform layer

Encord is a data-management and curation platform with native annotation for video, LiDAR, audio, text, and sensor fusion in one workflow, including SAM 2 natively ^[26].

Strengths: strong tooling for curating and annotating multimodal data you already have, with sensor-fusion support in a single workflow. A good fit if your bottleneck is organizing and labeling existing footage over collecting new data.

Limitations: Encord is a platform and tooling layer, so it won't collect first-person video for you. You bring the data; it helps you curate and annotate it. If your gap is supply, Encord alone won't fill it.

Best for: teams that already have egocentric footage and need a platform to curate, annotate, and manage it.

Others worth knowing

A few more names round out the field, listed here for completeness with sourcing noted. Luel sells rights-cleared egocentric datasets, and Awign and Lightwheel run large capture operations. These figures circulate via third-party rankings, not first-party verification, so confirm any number directly with the provider before relying on it. On the open side, Ego4D and Ego-Exo4D remain the canonical research references ^{[17] [12]}, and EgoVid-5M (over 5 million clips, presented at NeurIPS 2025) is the largest egocentric corpus aimed at video generation ^{[27] [28]}.

Build vs buy vs open dataset: when to use Ego4D or EgoDex vs a provider

The provider question only matters once you know you need custom capture, and often you don't yet. The open egocentric corpus is richer than most teams realize: EPIC-KITCHENS-100 alone is 100 hours with 90K action segments across 45 kitchens ^[18], EgoVid-5M brings over 5 million clips with kinematic-control and textual action annotations for generation work ^[27], and Egocentric-10K hands you 10,000 hours of factory-floor footage for free ^[5]. Start with what an open dataset can answer, then escalate to a vendor when the public corpora run out of fit.

• Use an open dataset when you are pretraining or running ablations and the scenes are close enough: Egocentric-10K for sheer factory-floor scale (10,000 hrs, Apache-2.0), EgoDex for dexterous tabletop tasks (829 hrs), EgoVid-5M for generation work.
• Buy custom capture when your embodiment, sensor stack, task distribution, or environment is not represented in any public corpus, or when you need exclusivity.
• Buy when you need rights you can actually ship on: research-licensed public data does not grant commercial-use rights for a deployed product.
• Buy when you need delivery in RLDS or LeRobot with provenance attached, instead of reformatting and re-clearing a public dataset yourself.

Can you use Ego4D or EgoDex commercially?

Check each license before you assume yes. Ego4D's access terms are research-oriented, so treating it as commercial training supply is a legal risk. EgoDex contains 829 hours across 194 tabletop manipulation tasks ^[13]. Apple's public release is CC BY-NC-ND—noncommercial and no derivatives—so it is not commercial data supply ^[29]. Egocentric-10K is explicitly Apache-2.0 ^[5], but you are still responsible for verifying the license text and any contributor-consent constraints for your specific commercial use. Truelabel does not distribute these public datasets. The safe default: read the license, and where you need cleared commercial rights with provenance, source custom data with rights attached instead of retrofitting a research corpus.

How much does egocentric video data cost?

There's no single price, and any provider quoting a flat per-hour rate without seeing your spec is guessing. Cost scales with four things, and you can read your own quote by which of them your spec demands. Enrichment depth comes first: hand and gaze tracking and action segmentation cost more than raw RGB, because each layer is human or model labor on top of the footage. Multimodal sync is second: time-aligned RGB-D plus IMU runs higher than video alone, since the alignment itself is engineering work. Exclusivity is third: non-shared data you own outright costs more than a slice of a shared corpus that other buyers also license. Environment difficulty is fourth: a controlled tabletop is cheaper than a busy real-world workplace where bystander consent, location releases, and retention terms all add overhead before a single frame is usable.

Separate the price of pixels from the price of fitness. Open corpora like Egocentric-10K are free under Apache-2.0 ^[5] and Ego4D is free for research ^[17], so if your scenes are close enough to the public data, your marginal cost is near zero. The moment your embodiment, sensors, or environment diverge, you're paying for capture, and the cost is set by how specific your spec is. A loose spec is cheap and produces footage you may not be able to use, while a tight spec with an acceptance rubric costs more per hour and wastes far less. A marketplace prices each spec against its enrichment, sync, exclusivity, and environment demands, so the most reliable way to read your own cost is to send a tight spec plus a sample request and price the matched batch.

What a good egocentric sourcing spec contains

Whether you buy from a marketplace or a managed service, the quality of what you get back is set by the quality of the spec you send. Vague briefs produce generic footage. Teams that get usable data define the request the way they'd define an evaluation: concretely, with acceptance criteria, before any capture starts. A strong spec names the embodiment and viewpoint, the task distribution, the modalities and their sync tolerance, the environment and diversity targets, the enrichment layers, the delivery format, and the licensing and consent requirements. Then it sets an accepted-trajectory target so a first batch can be graded against the rubric before scale-up. Specify, sample, then gate before you fund volume.

• Embodiment and viewpoint: human head-mounted, robot-mounted, ego-exo paired, or wrist-camera.
• Task distribution: the specific manipulations, navigations, or interactions your policy must learn.
• Modalities and sync: RGB, RGB-D, IMU, audio, hand-pose, gaze, and the alignment tolerance you require.
• Environment and diversity: regions, lighting, clutter, and workplace types that match deployment.
• Enrichment and delivery: required annotation layers and the target format (RLDS, LeRobot, Open X-Embodiment).
• Governance: commercial-use rights, contributor consent, location release, and retention terms.
• Acceptance: an accepted-trajectory target and QA threshold the first batch must clear before scale-up.

How to choose the right provider for your team

Match the provider to your stage and your binding constraint, not to the longest feature list.

• Research lab pretraining: start with open corpora (Egocentric-10K, EgoDex, Ego4D as a reference) before paying anyone.
• Scaling VLA startup needing deployment-fit data: a marketplace like Truelabel for custom first-person capture with robotics-ready delivery.
• Enterprise needing compliant volume and QA: iMerit or Scale AI for managed annotation, Shaip when consent and regional compliance bind.
• Already have footage, need to curate and label it: Encord for the platform layer, Objectways or Labellerr for managed annotation.
• Always run a first-batch eval against your own rubric before scaling up, whoever you pick.

The bottom line

No provider wins every dimension. For pretraining, start with open corpora. For deployment-fit first-person data with provenance and robotics-ready delivery, a marketplace model fits, which is where Truelabel leads. For managed enterprise volume, iMerit and Scale AI lead. Define your embodiment, environment, and accepted-trajectory targets up front, gate scale-up on a sample eval, and if you want custom egocentric capture routed to candidate collectors reviewed against the buyer spec with rights and metadata attached, post a spec.

Related pages

Use these to move from category-level context into specific task, dataset, format, and comparison detail.

External references and source context

1. truelabel egocentric data licensing hub

   Truelabel attaches consent artifacts, location releases, and per-session metadata sufficient for downstream license verification to egocentric capture.

   truelabel.ai ↩
2. iMerit Egocentric Video Data Collection Services

   iMerit provides dexterous-manipulation and video-data-collection services with 25,000+ domain experts across 60+ countries and SOC2/ISO 27001/GDPR/HIPAA compliance.

   imerit.ai ↩
3. scale.com scale ai universal robots physical ai

   Scale AI describes a Universal Robots collaboration for industrial physical AI.

   scale.com ↩
4. Appen Physical AI Data Annotation & Evaluation (case study)

   A frontier robotics lab partnered with Appen to annotate egocentric human video and evaluate robot performance, delivering 50,000+ units of Physical AI training data; egocentric annotation segmented videos into task intervals labeled with timestamp, task type, hand configuration, and a natural-language description.

   appen.com ↩
5. Egocentric-10K

   Egocentric-10K is 10,000 hours and 1.08 billion frames of real factory-worker first-person video at 30fps/1080p, released under Apache-2.0 with 192,900 clips.

   Hugging Face ↩
6. truelabel Open X-Embodiment glossary

   Open X-Embodiment aggregates 1M+ real robot trajectories across 22 embodiments and 527 skills from 60 datasets.

   truelabel.ai ↩
7. OpenVLA project

   OpenVLA is a 7B-parameter open-source vision-language-action model pretrained on 970k robot episodes from Open X-Embodiment.

   openvla.github.io ↩
8. Project Go-Big: Internet-Scale Humanoid Pretraining and Direct Human-to-Robot Transfer

   Figure trained its Helix navigation policy using 100% egocentric human video collected passively in real Brookfield homes, translating human navigation strategies into robot control with no robot demonstrations.

   Figure ↩
9. EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data

   NVIDIA EgoScale pretrained on over 20k hours of action-labeled egocentric human video and improved average task success rate by 54% over a no-pretraining baseline using a 22-DoF robotic hand.

   arXiv ↩
10. Project site

    DROID gathered 76k demonstration trajectories over 12 months across 564 scenes and three continents.

    droid-dataset.github.io ↩
11. EgoMimic: Scaling Imitation Learning via Egocentric Video

    EgoMimic co-trains manipulation policies from human egocentric video and teleoperated robot data using Project Aria glasses, and reports that scaling one hour of additional hand data is significantly more valuable than one hour of additional robot data.

    EgoMimic (Georgia Tech) ↩
12. Ego-Exo4D project site

    Ego-Exo4D is the largest public dataset of time-synchronized first- and third-person video.

    ego-exo4d-data.org ↩
13. EgoDex: Learning Dexterous Manipulation from Large-Scale Egocentric Video

    EgoDex contains 829 hours across 194 tabletop manipulation tasks, collected with Apple Vision Pro.

    arXiv ↩
14. EgoLive: A Large-Scale Egocentric Dataset from Real-World Human Tasks

    EgoLive provides 1,680 hours of stereo egocentric video at 60fps across 65,866 episodes spanning 346 real-world service tasks for robot manipulation learning.

    arXiv ↩
15. LeRobot documentation

    LeRobotDataset is a standard format robotics teams ingest robot-learning data into.

    Hugging Face ↩
16. MANUS RoboBrain-Dex: High-Fidelity Egocentric Data

    EgoAtlas, a multi-source egocentric dataset integrating human and robotic manipulation data under a unified action space, was constructed by researchers at Peking University and the Beijing Academy of Artificial Intelligence using MANUS Quantum Metagloves, which capture 3D positions for all 25 hand keypoints per hand.

    MANUS ↩
17. Egocentric video remains useful but incomplete for robot data buyers

    Ego4D is a massive-scale egocentric dataset with over 3,670 hours of daily-life video from 931 participants across 74 locations in 9 countries.

    ego4d-data.org ↩
18. EPIC-KITCHENS-100 dataset page

    EPIC-KITCHENS-100 is a 100-hour egocentric kitchen dataset with 90K action segments across 45 kitchens in 4 cities.

    epic-kitchens.github.io ↩
19. Egocentric video remains useful but incomplete for robot data buyers

    Ego4D is more than 20x greater than any other dataset in terms of hours of footage.

    ego4d-data.org ↩
20. HD-EPIC: A Highly-Detailed Egocentric Video Dataset

    HD-EPIC is a highly-detailed egocentric kitchen dataset of 41 hours across 9 kitchens with 59,000 fine-grained actions, all annotations grounded in 3D through digital twinning.

    arXiv ↩
21. Labellerr egocentric data services

    Labellerr offers first-person capture via wearable rigs and robot-mounted cameras with RGB and RGB-D support, a 5,000+ annotator workforce, and claims up to 99% annotation accuracy.

    labellerr.com ↩
22. Best VLA Training Data Providers (2026)

    Truelabel routes egocentric capture specs to candidate collectors for sample review, consent/provenance checks, buyer-approved acceptance gates, and requested delivery formats such as RLDS or LeRobot.

    truelabel.ai ↩
23. Appen Physical AI Training Data

    Appen lists Physical AI as a data-product area covering egocentric video collection and annotation, LiDAR annotation, robotics trajectories, sensor fusion, and robot-performance evaluation, sourced through a global contributor network of over 1 million.

    appen.com ↩
24. Objectways egocentric data collection

    Objectways delivers embodied-AI data via teleoperation, egocentric capture, and robot-trajectory annotation with 2,200+ trained annotators and SOC2 Type II / ISO 27001 / GDPR / HIPAA compliance.

    objectways.com ↩
25. Shaip managed data collection

    Shaip sources and curates datasets from over 60 countries with GDPR/HIPAA/ISO 9001/SOC 2 Type II/ISO 27001 compliance and multimodal robotics-and-autonomy data.

    shaip.com ↩
26. Encord data collection services

    Encord is a data-management and curation platform with native video, LiDAR, audio, text, and sensor-fusion annotation in one workflow, including SAM 2 natively.

    encord.com ↩
27. EgoVid-5M

    EgoVid-5M is over 5 million egocentric video clips built specifically for egocentric video generation, with kinematic and textual action annotations.

    EgoVid-5M project ↩
28. EgoVid-5M

    EgoVid-5M was presented as a NeurIPS 2025 poster.

    EgoVid-5M project ↩
29. EgoDex: code and dataset release

    Apple's public EgoDex release is CC BY-NC-ND: noncommercial and no derivatives, so it is not commercial data supply.

    Apple ↩

FAQ