Best robotics data providers in 2026: where to get ready-made robotics datasets
Answer in brief: Start with the free open datasets. FineART from Scale AI and Hugging Face is CC-BY-4.0, with 40,543 bimanual episodes and 1,718 hours. DROID has 76k trajectories across 564 scenes. Open X-Embodiment has 1M+ trajectories from 22 robot types. NVIDIA's Physical AI collection on Hugging Face adds simulated and teleoperation data. To license commercial data that already exists, go to Toloka (25,000+ hours of egocentric video, 500+ hours of UMI data and teleoperation data, with samples in 24 hours) or Firsthand (10,585 validated egocentric hours under a buyer-owned license). To generate your own data, use NVIDIA Isaac Sim (free, Apache 2.0) or Anyverse. If nothing off the shelf fits, Scale AI and Encord collect new data to your spec.
Last updated: 2026-11-09
Top pick: Scale AI. It is the only provider here that gives you both a large free dataset today (FineART: 533,913 human-written subtask labels in LeRobot format) and an enterprise data engine that Scale says receives more than 1,000 hours of new demonstrations every day. Read the Scale AI review ·
This guide answers one question: where do you get robotics data that already exists? That covers open datasets you can download, commercial catalogs you can license, simulators that generate data on demand, and the providers to call when you need new data. This is independent research. Every fact comes from the provider's own website or the dataset's official page, checked on the date above. We did not test any product. See our methodology, pricing comparison and glossary. If you are mainly hiring a team to collect or label new data, our sister guide on robotics data companies covers that in depth.
Start free: the open robotics datasets to download first
- FineART (Scale AI and Hugging Face LeRobot). 40,543 teleoperated episodes, 1,718 hours, 151 tasks and 533,913 subtask labels, all recorded on a Trossen Stationary AI (ALOHA-style) bimanual robot. LeRobot v3.0 format, 1.86 TB, CC-BY-4.0. lerobot/FineART
- Open X-Embodiment. 1M+ real robot trajectories across 22 embodiments, pooled from 60 datasets from 34 labs, in RLDS episode format. The repository says its software is Apache 2.0 and its other materials are CC-BY 4.0, and asks you to cite each contributing dataset. robotics-transformer-x.github.io
- DROID. 76k demonstration trajectories (350 hours) across 564 scenes and 86 tasks, recorded on Franka Panda arms. It loads with TensorFlow Datasets. Language annotations (December 2024) and improved camera calibrations (April 2025) were added after release. droid-dataset.github.io
- NVIDIA Physical AI collection. NVIDIA calls these open, commercial-grade datasets. They include GR00T teleoperation, kitchen manipulation, GraspGen, SimReady warehouse scenes, and BridgeData2 and LIBERO converted to LeRobot v3. huggingface.co/collections/nvidia/physical-ai
- Toloka HomER. A small open egocentric dataset: 63 videos across 17 household task categories. Toloka says it took 3 hours and $50 to collect. toloka/HomER
- Perception and driving datasets. Segments.ai maintains a directory of open datasets, including nuScenes, Waymo, Argoverse 2, KITTI, SemanticKITTI and ScanNet. segments.ai/resources/open-datasets
For general-purpose AI datasets beyond robotics, see our sister site best AI datasets.
The ranked list
1. Scale AI: Physical AI Data Engine
Ready-made data: FineART, the free CC-BY-4.0 dataset described above. Custom datasets are sold under enterprise contract.
What it is: A service that collects, labels, validates and delivers real-world robotics demonstrations. You can collect three ways: teleoperated bimanual robots, Scale Harness (a first-person rig that needs no robot), or your own hardware, managed and operated by Scale. Collection runs in data factories, in collectors' homes, and at commercial and industrial sites.
Best for: Teams building robot foundation models or VLA models that need high volume, many embodiments and action-level labels.
Strengths: Scale says more than 1,000 hours of demonstrations reach its platform every day, and that datasets go through internal policy fine-tuning and evaluation before they are listed. It names Generalist and Physical Intelligence as partners. It says it is SOC 2 Type II and ISO 27001 certified and supports GDPR, CCPA and optional onshore data processing. FineART was released alongside FineART-VLA, an open model in the LeRobot ecosystem. About 23% of FineART's hours cover contact-rich skills such as deformable objects, tool use, cable routing and precision insertion.
Drawbacks: The engine has no public pricing, license terms or catalog. FineART covers a single robot type. The volume figures are Scale's own.
Pricing model: FineART is free under CC-BY-4.0. Engine data is an enterprise contract; book a demo.
2. Encord: robotics data collection plus annotation platform
Ready-made data: E-MM1, an open multimodal dataset of 107M data groups across images, video, audio, text and 3D point clouds. It is not robot-action data.
What it is: A multimodal data platform plus managed data services. Encord says it collects teleoperation, egocentric and embodied data in its own warehouse facilities and adapts the collection pipeline to your hardware and task protocols. Its platform labels synchronized RGB, depth, LiDAR and force/torque data in one interface.
Best for: Teams that want one vendor for the whole cycle: collect, curate, label, then train.
Strengths: 6-DoF pose, keypoint and hand-object labels for grasps. Supported data types include UMI, stereo video, RGB-D, LiDAR, force sequences, teleoperation recordings and egocentric video. Embedding-based curation finds rare scenarios before training. Encord says your data stays in your cloud, and that it has completed a SOC 2 examination and is HIPAA and GDPR compliant. Customer stories it cites include a 30% annotation-accuracy gain for Pickle Robot and a 70% productivity increase for Archetype AI.
Drawbacks: No licensable robot-action dataset. The pricing page lists tiers without dollar amounts, and 3D, LiDAR and point-cloud labeling is an add-on. The cited results come from its own case studies.
Pricing model: Starter, Team and Enterprise platform tiers (Enterprise is contact sales). Collection and managed services are scoped per project.
3. Toloka: licensable physical-AI datasets and custom data programs
Ready-made data: The largest licensable physical-AI catalog in this guide.
What it is: A data company that supplies demonstrations, annotation and evaluation for physical AI, and licenses an off-the-shelf dataset catalog.
Best for: Teams that want existing teleoperation, manipulation or egocentric data quickly, without waiting for a custom collection.
Strengths: Egocentric household video: 25,000+ hours and 216,000+ clips across 12 activity categories and 24 countries, with 7 time-aligned annotation layers. UMI manipulation: 500+ hours from a head camera and two handheld gripper cameras (1080p30), a 100 Hz IMU, 6-DoF pose and millimeter gripper aperture. Teleoperation from VR-driven bimanual and single-arm robots, with 25 Hz state and actions and wrist and head RGB-D cameras. Exocentric (third-person) video. Samples arrive within 24 hours, commercial training use is allowed, and production pipelines can go live in 48 hours. A self-serve egocentric collection preset has no minimums.
Drawbacks: No published prices or license terms. A fixed catalog may not match your robot. Much of Toloka's catalog is LLM and agent data.
Pricing model: Dataset license or custom program; contact Toloka.
4. Kognic: multi-sensor annotation for autonomy
Ready-made data: None. Kognic labels sensor logs you already have.
What it is: An annotation platform plus a specialist workforce, founded in 2018 in Gothenburg, Sweden. It labels camera, LiDAR and radar data for perception, planning and end-to-end models.
Best for: Autonomous-vehicle, ADAS and mobile-robot teams that need production-grade labels on their own calibrated multi-sensor logs.
Strengths: Kognic says it has delivered more than 100 million annotations across more than 120 projects. Listed customers include Qualcomm, BMW, Zenseact, AUMOVIO, Bosch, Kodiak, ZF, Einride and JLR. It runs more than 90 automated quality checkers and offers Language Grounding for VLA models. It is ISO/IEC 27001:2022 and SOC 2 Type II certified and TISAX assessed. You can buy it as a managed service, a platform license or a hybrid.
Drawbacks: Built mainly for autonomous driving. No public price list or freemium tier.
Pricing model: Enterprise, per project.
5. Firsthand: custom and catalog egocentric datasets
Ready-made data: A catalog of 10,585 validated hours across six environments, plus a free 2.4 GB sample pack.
What it is: Collects first-person video for embodied AI. On its instrumented rig, each episode has ten synchronized streams: a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, segmentation and action labels.
Best for: Teams that need human demonstration video with rich signals, documented consent and a commercial license the buyer keeps.
Strengths: Reports a median cross-stream sync error of 1.12 ms on its instrumented rig. Delivers in RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. The license is perpetual and buyer-owned, with exclusivity available. It reports a 48-hour median time to first delivery (its own figure).
Drawbacks: Human video only, with no robot teleoperation. Thin areas such as clinical, agriculture and hospitality take three to four weeks to first batch.
Pricing model: Billed per validated hour; quote on request.
6. Guardian AI: verified egocentric video plus teleoperation
Ready-made data: No public catalog. Data is collected to order.
What it is: Collects first-person training data across Asia and the US. Its operators also drive customers' robots remotely to record demonstrations.
Best for: Teams that need proof of where data came from and who consented, or that want teleoperation without staffing their own operators.
Strengths: Four provenance tiers, P0 to P3. At the top tier, clips are signed on the recording device and screened for synthetic or re-filmed footage. Labels include 3D hand pose, 6-DoF camera trajectory, hand-object contact, task segmentation and success, failure and recovery outcomes. Delivers in LeRobot, RLDS, WebDataset, HDF5 and MCAP.
Drawbacks: A younger, smaller company. Engagements start with a paid pilot of at least 20 finished hours per week.
Pricing model: Paid pilot, then a price per finished hour.
7. NVIDIA Isaac Sim, Isaac Lab and Cosmos: simulation and synthetic data
Ready-made data: The Physical AI open datasets on Hugging Face, plus as much simulated data as you want to generate yourself.
What it is: NVIDIA's robotics simulation stack. Isaac Sim is an open-source framework for simulation, testing and synthetic data generation. Isaac Lab trains robot policies, with 16+ robot models and 30+ ready-to-train environments. Cosmos world foundation models augment data. Isaac Capture records demonstrations.
Best for: Teams that want to generate their own data in-house for perception, manipulation, mobility, reinforcement learning and testing.
Strengths: Physics and sensor simulation with domain randomization. Built-in annotators produce RGB, bounding boxes and instance and semantic segmentation, with COCO and KITTI export. Imports URDF, MJCF and CAD, bridges to ROS 2, and runs on Brev, NGC or AWS Marketplace.
Drawbacks: It is software, so you need simulation skills and NVIDIA GPUs. Simulated data still has to be validated against real-world performance. Licensing is layered: the Isaac Sim source is Apache 2.0, but Omniverse Kit and bundled assets fall under separate NVIDIA terms.
Pricing model: NVIDIA says Isaac Sim is free to use. Redistributing Omniverse Kit in your own product needs a separate license, which NVIDIA's developer FAQ says is available through an Omniverse Enterprise subscription. For reference, NVIDIA AI Enterprise list prices are $4,500 per GPU for one year, $22,500 per GPU perpetual with 5 years of support, or $1 per GPU-hour on cloud marketplaces plus instance costs. Isaac Lab is BSD-3. GPU and cloud costs are extra.
8. micro1: expert-demonstration robotics data
Ready-made data: No public catalog. It turns raw robotics data, including data you already hold, into structured training data.
What it is: A human-data company. It records first-person and third-person video of domain experts doing real tasks, structures and labels raw robotics data, and runs a full-stack deployment lab where it operates robots and captures interaction data.
Best for: Demonstrations of skilled work such as repair, assembly, logistics and field operations, and turning raw logs into training-ready datasets.
Strengths: Multimodal capture pipelines with synchronized stereo video, IMU and multi-camera data. Dense action-level annotations: temporal segmentation, hand-object interactions and step-by-step natural-language labels for VLA training and evaluation. Labeling, tracking, 2D and 3D annotation, QC and evaluation in one service. Its in-house robot lab connects data capture to end-to-end evaluation.
Drawbacks: No public dataset catalog, samples or pricing. Its robotics page does not list teleoperation.
Pricing model: Custom engagement; contact micro1.
9. Segments.ai: point-cloud and sensor-fusion labeling platform
Ready-made data: A free directory of open perception datasets you can clone and adapt.
What it is: Labeling software for images, point clouds, sequences and multi-sensor setups. If you need people to do the labeling, it connects you to partner labeling companies.
Best for: Robotics teams that label 3D data in-house and want clearer pricing than most vendors offer.
Strengths: Unlimited seats, projects and API/SDK access. ML-assisted labeling and active learning. A 14-day free trial and free academic licenses.
Drawbacks: Annual usage minimums apply. Labeling services come from partners, and Segments.ai does not collect data.
Pricing model: Every plan is a custom quote. Core starts at 3,600 labeling hours per year (stated as equivalent to $2.67 per hour). Fusion starts at 5,000 hours per year. Enterprise includes 150,000+ hours per year.
10. Sama: managed 3D and AV annotation
Ready-made data: Sama publishes a Sama-Coco dataset resource. Its core business is labeling your data.
What it is: A managed, ML-assisted annotation and model-validation service for driving and other sensor-heavy autonomy data. Sama also has a robotics and manufacturing practice.
Best for: Programs with many sensors and long sequences that want to hand off labeling and validation completely.
Strengths: Loads up to 14 sensors and multiple point clouds. Supports SLAM preprocessing, accumulated point clouds, global-coordinate conversion, and shutter and ego-motion compensation. Its SamaAssure program puts quality in writing: 95% as standard, up to 99.5%. Integrates through API, CLI and webhooks.
Drawbacks: Mainly focused on AV and ADAS. No rate card.
Pricing model: Custom per project.
11. Robata AI: expert demonstrations and humanoid teleoperation
Ready-made data: None published. Data is collected to order.
What it is: A provider of expert-demonstrated data for robotics labs. It records first-person and third-person video of industry experts at work, and runs multi-view teleoperation sessions on humanoids it operates.
Best for: Labs that need real workplace demonstrations and clean manipulation trajectories for imitation learning and VLA training.
Strengths: Both human video and robot teleoperation, captured with consistent hardware and QC. Fine-grained, multi-stream action labels come with natural-language descriptions, in JSONL or CSV.
Drawbacks: A very small public footprint: a one-page site with email contact, and no published volumes, sample data or pricing.
Pricing model: Custom quote by email.
12. Anyverse: synthetic multi-sensor perception data
Ready-made data: Synthetic perception datasets, generated on demand to your taxonomy.
What it is: A physics-based synthetic data and sensor-simulation platform for robot perception, aimed at autonomous mobile robots, forklifts and fleets.
Best for: Validating and training perception on camera, LiDAR, radar and thermal sensors, especially for rare, dangerous and safety-case scenarios.
Strengths: Deterministic rendering: a scenario renders identically on every run. One scene produces aligned ground truth for every sensor, including semantic and instance segmentation, depth, 3D position, normals, motion vectors and material. Procedural generation and distribution-based domain randomization. TISAX certified.
Drawbacks: Perception only, with no manipulation or action demonstrations. No public pricing.
Pricing model: Contact sales or book a demo.
Comparison table
| Rank | Provider | Ready-made data you can get | Main data type | Public pricing |
|---|---|---|---|---|
| 1 | Scale AI | FineART (free, CC-BY-4.0) | Teleoperation and egocentric demonstrations | No (FineART free) |
| 2 | Encord | E-MM1 open multimodal dataset | Facility collection plus multi-sensor annotation | Tiers listed, no prices |
| 3 | Toloka | Licensed egocentric, UMI, teleoperation and exocentric data; HomER (open) | Licensed datasets plus programs | No |
| 4 | Kognic | None; labels your logs | Camera/LiDAR/radar annotation | No |
| 5 | Firsthand | 10,585-hour egocentric catalog; free sample pack | Egocentric video, 10 streams | Per validated hour (quote) |
| 6 | Guardian AI | Collected to order | Verified egocentric video plus teleoperation | Per finished hour (quote) |
| 7 | NVIDIA Isaac Sim | Physical AI open datasets; self-generated synthetic data | Simulation and synthetic data | Free; enterprise list prices published |
| 8 | micro1 | Collected to order; structures your raw data | Expert demonstrations, action labels | No |
| 9 | Segments.ai | Directory of open perception datasets | Point-cloud and sensor-fusion labeling | Usage minimums published |
| 10 | Sama | Sama-Coco resource | Multi-sensor AV annotation and validation | No (quality guarantee published) |
| 11 | Robata AI | Collected to order | Expert demonstrations, humanoid teleoperation | No |
| 12 | Anyverse | Synthetic datasets on demand | Synthetic multi-sensor perception | No |
How to get the right robotics data fastest
- Use the open datasets first. FineART, DROID, Open X-Embodiment and NVIDIA's Physical AI collection are free. Pre-train or co-train on them before you pay for anything.
- Read every license. FineART is CC-BY-4.0. Open X-Embodiment states CC-BY 4.0 for its materials but asks you to cite each contributing dataset. Licenses on other open datasets vary.
- License a catalog to fill gaps. Toloka and Firsthand sell data that already exists, so you can review samples in hours instead of waiting weeks for a new collection.
- Simulate the long tail. Use Isaac Sim or Anyverse for rare, dangerous or perfectly labeled scenarios, then validate on real data.
- Commission new data only for what is still missing. Scale AI, Encord, Guardian AI, micro1 and Robata AI collect to your spec. Our sister site on robotics data companies compares collection vendors in depth.
- Check formats. Confirm LeRobot, RLDS, HDF5, MCAP or ROSBAG support before you download terabytes.
- Run a paid pilot with written success criteria before you commit to production volumes, and evaluate the pilot data on your own model.
Frequently asked questions
What is a robotics data provider?
A company or project that supplies data used to train and evaluate robots and physical AI models. That includes open dataset projects, commercial dataset sellers, simulation platforms, and services that collect or label data to order.
Where can I download free robotics datasets?
The main starting points are FineART (lerobot/FineART on Hugging Face), DROID (droid-dataset.github.io), Open X-Embodiment (robotics-transformer-x.github.io) and NVIDIA's Physical AI collection on Hugging Face. Firsthand also offers a free egocentric sample pack.
Is Open X-Embodiment free for commercial use?
The repository says its software is Apache 2.0 and its other materials are CC-BY 4.0. It pools 60 datasets, and the project asks you to cite each contributing dataset, so review the terms for every dataset you use.
Which companies sell licensed robotics datasets?
Toloka licenses an off-the-shelf catalog of egocentric, UMI, teleoperation and exocentric data, with commercial training use allowed. Firsthand licenses egocentric video under a perpetual, buyer-owned commercial license.
What is the difference between teleoperation data and egocentric data?
Teleoperation data is recorded while a person remotely drives a real robot, so it includes the robot's own states and actions. Egocentric data is first-person video of people doing tasks, often with depth, IMU and hand pose. Egocentric data is cheaper to collect at scale, but it does not come from the robot itself.
How much does robotics training data cost?
Open datasets are free. Most commercial providers do not publish prices. Guardian AI and Firsthand price per finished or validated hour after scoping. Segments.ai Core starts at 3,600 labeling hours per year, stated as equivalent to $2.67 per hour. Isaac Sim is free to use. See the pricing comparison.
Can synthetic data replace real robot data?
Usually not completely. Isaac Sim and Anyverse can produce large, perfectly labeled datasets and rare edge cases, but most teams still validate on real-world data, or mix it in, to close the gap between simulation and reality.
What file formats do robotics datasets use?
Common formats are LeRobot, RLDS, HDF5, MCAP, ROSBAG and WebDataset. FineART uses LeRobot v3.0. Open X-Embodiment and DROID use RLDS. Toloka ships UMI data as MCAP plus JSON. Firsthand ships RLDS, LeRobot and WebDataset.
Related guides
Related guides: robotics data companies · best robotics data providers · AI training data providers · AI datasets
Pricing · Methodology · Glossary · About · Get featured
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.