Robotics data companies: the 2026 directory by category
Answer in brief: Robotics data companies fall into five categories. Teleoperation collection: Scale AI, Labelbox Terra, Encord and Fax Research record robot demonstrations. Egocentric video: Firsthand, Toloka, Knoxlabs and Appen capture first-person human task footage. Sensor and LiDAR annotation: iMerit, Sama, Segments.ai and Kognic label camera, LiDAR and radar logs. Dataset marketplaces: Toloka, Fax Research and Appen license existing data, and NVIDIA publishes free open datasets. Simulation and synthetic data: NVIDIA Isaac Sim and Parallel Domain. For a fully managed program that covers several categories, start with Scale AI, Labelbox Terra or Encord.
Last updated: 2026-11-09
This is an independent research directory. We did not test these products or buy their data. Every entry is based on the company's own public pages, read on the date above. Figures such as hours, volumes and certifications are the vendors' own claims. See our methodology and the robotics data pricing guide.
Top pick for 2026: Scale AI Physical AI Data Engine. It is the broadest managed option in the directory. It collects teleoperated bimanual robot demonstrations, data from its robotless Scale Harness egocentric rig, and data from your own hardware, which Scale operates for you. Scale says more than 1,000 hours of demonstration data reach its platform every day. Data comes from data factories, home collectors and in-field deployments, and every demonstration gets grounding annotations. · Read the Scale AI review
The five categories of the robotics data industry
- Teleoperation collection: People drive real robots to record demonstrations with joint states and actions, used for post-training and imitation learning. Listed: Scale AI, Labelbox Terra, Encord, Fax Research, and Knoxlabs (pilot scale).
- Egocentric video: First-person footage of people doing tasks, often with wrist cameras, depth, pose or handheld-gripper (UMI) signals, used for pre-training. Listed: Firsthand, Toloka, Knoxlabs, Appen, plus the egocentric rigs at Scale AI, Labelbox Terra and Fax Research.
- Sensor and LiDAR annotation: 2D/3D labeling of your own camera, LiDAR, radar and IMU logs for perception. Listed: iMerit, Sama, Segments.ai, Kognic, and Encord and Appen.
- Dataset marketplaces: License data that already exists. Listed: Toloka's off-the-shelf catalog, Fax Research dataset sourcing and Appen's off-the-shelf sets. NVIDIA publishes free Physical AI Open Datasets on Hugging Face, including GR00T teleoperation and manipulation sets. For general AI datasets, see our sibling bestaidatasets.com.
- Simulation and synthetic data: Generated scenes and sensor data for scale, edge cases and testing. Listed: NVIDIA Isaac Sim (open source under Apache 2.0, built on Omniverse, exports annotations in COCO and KITTI formats) and Parallel Domain (PD Replica digital twins of scanned real facilities, plus PD Sim deterministic sensor simulation for warehouse robotics).
The directory at a glance
| Company | Primary category | Also covers | Review |
|---|---|---|---|
| Scale AI | Teleoperation collection | Egocentric video, annotation | Scale AI review |
| Labelbox Terra | Teleoperation collection | Egocentric video, evals | Labelbox review |
| Encord | Teleoperation collection | LiDAR annotation, curation software | Encord review |
| Toloka | Dataset marketplace | Self-serve egocentric collection | Toloka review |
| Appen | Egocentric video | LiDAR annotation, off-the-shelf sets | Appen review |
| Knoxlabs | Egocentric video | UMI/glove, teleop (pilot scale) | Knoxlabs review |
| Firsthand | Egocentric video | Custom collection | Firsthand review |
| Fax Research | Teleoperation collection | UMI, dataset sourcing, verification | Fax Research review |
| iMerit | Sensor and LiDAR annotation | Managed workforce | iMerit review |
| Sama | Sensor and LiDAR annotation | Model validation | Sama review |
| Segments.ai | Sensor and LiDAR annotation | Labeling software | Segments.ai review |
| Kognic | Sensor and LiDAR annotation | Language grounding for VLA | Kognic review |
The ranked list
Scale AI: Physical AI Data Engine
Category: Teleoperation collection · Egocentric video
What it is: A managed service that collects and annotates real-world robotics data. Collection runs through robotics data factories, at-home collectors around the world, and on-site deployments in commercial and industrial settings. Hardware options are teleoperated bimanual robots, a robotless egocentric rig called Scale Harness, and customer hardware that Scale operates for you.
Best for: Robot foundation-model and humanoid teams that need a large, steady volume of data and a fully managed program.
Standout strengths: Scale says more than 1,000 hours of demonstration data are uploaded to its platform every day. In-house roboticists write the data specs for each program. Scale says datasets go through its own policy fine-tuning and evaluation before they are listed. It names Generalist and Physical Intelligence as labs its platform supports, says it is SOC 2 Type II and ISO 27001 certified, and offers optional onshore data processing.
Honest drawbacks: No public pricing; access starts with a demo. Volume and model-improvement claims are self-reported. The white-glove enterprise model may be more than small labs need.
Pricing model: Custom, sales-led.
Labelbox: Terra
Category: Teleoperation collection · Egocentric video
What it is: Labelbox's robotics data engine. It captures egocentric and teleoperated demonstrations on purpose-built rigs, then validates and annotates them. Pre-training, post-training and evaluation data all use one episode format.
Best for: Teams that want commissioned demonstrations with measured coverage across environments, objects and tasks, plus standardized policy evaluations.
Standout strengths: Labelbox reports more than 1 PB of data produced as of January 2026. Terra has seven rig configurations, from a single head camera to a trio-plus-instrumented-gripper rig and a teleoperation rig, with up to four synced camera streams per episode. A diversity engine checks each episode against a target distribution and steers collection toward coverage gaps. Episodes that fail validation are re-recorded. Terra is embodiment-agnostic.
Honest drawbacks: Collection runs through Labelbox's Alignerr network and third-party partners, not only Labelbox staff. No public prices.
Pricing model: Each program is scoped and quoted from your task list and target distribution.
Encord
Category: Teleoperation collection · Sensor and LiDAR annotation
What it is: A data platform plus data-as-a-service for robotics, autonomous systems and embodied AI. It covers four phases: collection, curation, LiDAR/point-cloud annotation, and feedback after deployment.
Best for: Teams that want one vendor for both the software (curation and annotation, possibly in their own cloud) and managed collection.
Standout strengths: In-field operators and reconfigurable lab sets with teleoperation arms. Collection protocols are piloted at Encord facilities and can run on-site at your factory or warehouse. Native MCAP and ROSBAG support, plus multi-camera video, LiDAR, IMU and time-series data. VLA tooling includes timestamped action captioning, multi-step subtask labels and human-reviewed VLM pre-captions. Deployment failures captured through remote teleoperation feed back into collection. VPC, on-premise and air-gapped setups are available. Encord says it is SOC 2 Type II, HIPAA and GDPR certified and that 300+ AI teams use it.
Honest drawbacks: Its pricing page lists 3D, LiDAR and point-cloud annotation as an add-on and shows no dollar amounts. Its claim to be the only end-to-end Physical AI data provider is marketing; other vendors here offer similar scope.
Pricing model: Starter, Team and Enterprise platform tiers (prices not published); managed data services are custom.
Toloka
Category: Dataset marketplace · Egocentric video
What it is: An AI data company with an off-the-shelf catalog that includes physical AI datasets (egocentric video, UMI manipulation, teleoperation, exocentric video), custom data programs, and a self-serve egocentric video collection platform.
Best for: Teams that want to license existing robotics data now, or launch a quick egocentric collection without a sales process.
Standout strengths: Toloka reports 25,000+ hours of physical AI data. Its egocentric household dataset lists 216,000+ clips across 12 activity categories and 24 countries, with seven time-aligned annotation layers. The UMI dataset (500+ hours) records a head camera plus two hand-held gripper cameras at 1080p30, a 100 Hz IMU, 6-DoF pose and gripper aperture in millimeters. The teleoperation data covers VR-driven bimanual and single-arm robots, with 25 Hz state and actions and wrist and head RGB-D. Samples arrive within 24 hours. The self-serve platform has no sales cycle and no minimums; Toloka says it has 200,000+ contributors in 100+ countries and delivers a first batch in about 3 hours. Its example open dataset, HomER, is on Hugging Face (63 videos across 17 household task categories); Toloka says it took 3 hours to collect and cost $50.
Honest drawbacks: Much of the catalog targets LLM and agent training rather than robotics. Licensing terms and self-serve rates are not published. Self-serve QA is LLM-based; Toloka reports 89.1% accuracy at catching failures, so some bad clips can get through.
Pricing model: Dataset licensing or custom programs through sales; self-serve collection is pay-as-you-go (rates not published).
Appen
Category: Egocentric video · Sensor and LiDAR annotation · Dataset marketplace
What it is: A long-established AI training data company with a Physical AI line. It offers stereo egocentric capture with synchronized wrist pose and gripper state, human task demonstrations, camera/LiDAR/radar sensor fusion data, LiDAR annotation, autonomous-vehicle data and world-model data.
Best for: Enterprises that need documented quality measurement, consent and provenance records, and controlled capture facilities.
Standout strengths: Hardware-synchronized rigs record to a shared clock and ship with sub-millisecond alignment metadata and intrinsic/extrinsic calibration. On-site kitchen, workshop, assembly and logistics environments. Appen says it has a contributor network of 1M+ people in 170+ countries. Quality controls include inter-annotator agreement standards, multi-pass review and gold-standard task validation. Off-the-shelf sets include 82,000 robot-vacuum-perspective images and 1.2M CAD files.
Honest drawbacks: Physical AI covers six service lines in a broad portfolio. Its Physical AI page names no robot embodiments or delivery formats and does not describe robot teleoperation. Its off-the-shelf physical AI sets are mostly perception imagery, not manipulation demonstrations. No public rate card.
Pricing model: Custom quote after scoping.
Knoxlabs
Category: Egocentric video · Teleoperation collection (pilot scale)
What it is: A Glendale, California company that collects human demonstration and teleoperation data in real US homes and at its Los Angeles facility. Capture modes are egocentric, UMI/wearable glove, dexterous hand and robot teleop.
Best for: Teams that want to start small (sample pack, then pilot) with US home-environment data, and teams already buying gloves or capture hardware.
Standout strengths: Knoxlabs says it deployed and fleet-managed a capture fleet of about 2,000 headsets that collected 10M+ episodes in US homes. Delivery is in HDF5, RLDS/LeRobot or your own spec. Sample episodes and schema docs arrive within one business day. It can activate existing VR headset owners without shipping hardware, and it sources capture hardware such as MANUS, SenseGlove and Xsens. Quoted hours are post-QA delivered hours.
Honest drawbacks: Its own page says dexterous-hand and teleop collection is at pilot scale, expanding with the LA facility in Q4 2026. Collection is US-focused. No public prices.
Pricing model: Free samples, then a scoped pilot, then custom pricing at scale.
Firsthand
Category: Egocentric video
What it is: A specialist in custom egocentric video datasets for embodied AI and robotics, collected to the buyer's spec.
Best for: Teams pre-training on rich first-person human data who want to inspect real data before talking to sales.
Standout strengths: Each episode from its instrumented rig ships ten synchronized streams, including a 4K head camera, two wrist cameras, depth, IMU, 3D hand and body pose, segmentation and action labels. Firsthand reports a median cross-stream sync error of 1.12 ms, 10,585 validated hours delivered, and a median 48 hours to first delivery. Formats include RLDS, LeRobot and WebDataset. Consent is documented and the buyer owns a perpetual commercial license. A free sample pack of about 2.4 GB is available without a call or NDA.
Honest drawbacks: Human egocentric data only; no robot teleoperation trajectories. A second capture tier using contributor-worn 360 cameras has no depth, no wrist views and no sub-millisecond sync. Its global delivery case study is anonymized. Per-hour rates are not published.
Pricing model: Billed per validated hour (rate on request).
Fax Research
Category: Teleoperation collection · Dataset marketplace
What it is: A physical AI data company offering egocentric video, UMI handheld-gripper demonstrations, hardware-specific robot teleoperation, annotation, and data verification.
Best for: Buyers who care most about provenance, meaning proof of where data came from and that it is authentic, especially when licensing existing footage.
Standout strengths: Offers both off-the-shelf licensed datasets and custom task collection. Verification covers source and rights, file and signal integrity, content authenticity and training suitability. Deliverables may include QC findings, data manifests and review notes. Member of NVIDIA Inception.
Honest drawbacks: Little public evidence of delivered volume or customers. It states that verification is not an absolute guarantee against synthetic or manipulated data. No public prices.
Pricing model: Custom, scoped per dataset or program.
iMerit
Category: Sensor and LiDAR annotation
What it is: A managed data annotation company with a robotics practice covering perception for household, logistics, warehouse, agriculture, medical, delivery and aerial robots.
Best for: Robotics perception teams that need large-scale 2D/3D labeling of their own sensor data.
Standout strengths: iMerit reports 6,000+ annotators and 10+ delivery centers. It covers multi-sensor fusion, 3D point clouds, panoptic and semantic segmentation, and object tracking. It lists ISO 27001, SOC 2, GDPR and HIPAA compliance. Its Carbon Robotics case study describes nearly 1 million images annotated for precision weeding.
Honest drawbacks: Focused on perception labeling; its robotics page does not describe teleoperation or demonstration collection. No public pricing.
Pricing model: Custom quote.
Sama
Category: Sensor and LiDAR annotation
What it is: A data annotation and model validation company with a robotics and manufacturing offering covering autonomous mobile robots, cobots, palletization, quality control and package inspection.
Best for: Industrial and warehouse robotics teams that want a written quality guarantee on 2D/3D annotation.
Standout strengths: Loads up to 14 sensors (cameras, LiDAR, radar, ultrasonic), with SLAM, accumulated point clouds and ego-motion compensation. SamaAssure starts at a 95% written quality guarantee and can go up to 99.5%. Sama also offers model validation and has been recertified as a B Corp.
Honest drawbacks: Annotation and validation only; its robotics page offers no demonstration or teleoperation collection. No public rate card.
Pricing model: Project-scoped quote.
Segments.ai
Category: Sensor and LiDAR annotation (software)
What it is: A multi-sensor 2D/3D labeling platform for robotics and autonomous-vehicle data. You run your own labeling on it; for labeling services it connects you with partner vendors.
Best for: Teams with their own labelers, or a partner vendor, who want point-cloud and sensor-fusion tooling with fairly clear pricing.
Standout strengths: Track IDs stay in sync across sensors and over time, and labels project from 3D to 2D. Cuboid tools, merged point clouds, keyframe interpolation and ML-assisted labeling. The Core plan starts at 3,600 labeling hours a year (Segments.ai says this works out to $2.67/hour) with a 14-day free trial. Academic licenses are free.
Honest drawbacks: No data collection. Sensor fusion interfaces and unlimited point-cloud size require the quote-based Fusion plan; Core caps point clouds at 500,000 points.
Pricing model: Annual usage-based quotes (Core, Fusion, Enterprise).
Kognic
Category: Sensor and LiDAR annotation
What it is: An annotation platform and workforce service founded in 2018 in Gothenburg, Sweden. Its homepage describes it as purpose-built for autonomous driving; it also says it works with robotics and industrial automation companies.
Best for: Teams with camera, LiDAR and radar data, especially vehicle-like mobile robots, who want a flexible mix of in-house and outsourced annotation.
Standout strengths: Three ways to buy: managed service, platform license or hybrid. Kognic reports 100M+ annotations across 120+ projects and names customers including Qualcomm, BMW, Zenseact, AUMOVIO and Bosch. It offers native 3D/LiDAR annotation with sensor fusion, more than 90 automated quality checkers, and Language Grounding workflows for vision-language-action models. It says it is ISO/IEC 27001:2022 and SOC 2 Type II certified and TISAX assessed.
Honest drawbacks: Product and case studies are driving-first, with little public evidence of manipulation or humanoid work. No public prices.
Pricing model: Custom, after contact and evaluation.
Comparison table
| # | Company | Type | Collects new data | Teleop data | Egocentric / UMI | 3D / LiDAR labeling | Off-the-shelf datasets | Public pricing |
|---|---|---|---|---|---|---|---|---|
| 1 | Scale AI | Managed data engine | Yes | Yes | Yes | Annotation included | Listed datasets | No |
| 2 | Labelbox Terra | Managed data engine | Yes | Yes | Yes | Not stated for Terra | Corpus samples | No |
| 3 | Encord | Platform + managed service | Yes | Yes | Yes | Yes (add-on) | No | Tiers, no prices |
| 4 | Toloka | Dataset licensing, custom + self-serve | Yes | Yes (dataset) | Yes | Not stated | Yes | No |
| 5 | Appen | Managed service | Yes | Not stated | Egocentric | Yes | Perception sets | No |
| 6 | Knoxlabs | Collection specialist | Yes | Pilot scale | Yes | No | Samples | No |
| 7 | Firsthand | Egocentric specialist | Yes | No | Egocentric | No | Free sample pack | Per-hour model, rate on request |
| 8 | Fax Research | Collection + verification | Yes | Yes | Yes | Not stated | Yes (licensed) | No |
| 9 | iMerit | Managed annotation | No | No | No | Yes | No | No |
| 10 | Sama | Managed annotation + validation | No | No | No | Yes | No | No |
| 11 | Segments.ai | Labeling software | No | No | No | Yes | No | Partial ($2.67/hr equivalent on Core) |
| 12 | Kognic | Annotation platform + service (driving-first) | No | No | No | Yes | No | No |
Not stated means the company's public page did not say either way.
Tools often used alongside these vendors
These are not data suppliers, but robotics teams use them to manage, view or generate data. Each description is based on the tool's own site, read 2026-11-09.
- Foxglove: an agentic data platform for robotics. It records logs in the open-source MCAP format, uploads ROS 1 bag and MCAP files from robots, visualizes recordings, connects to live robots for visualization or teleoperation, and searches and curates data across a fleet. Foxglove says it is SOC 2 Type II compliant. It runs in Foxglove Cloud, in your own storage bucket or on-premises, and you can get started for free.
- Rerun: an open-source data layer for Physical AI. It logs, stores and converts multimodal data in .rrd files, with dataframe and SQL queries. Rerun Hub is its production backend for cataloging recordings and streaming dataset mixes to training.
- Voxel51: search, visualization, curation and labeling for multimodal physical AI data, with synchronized playback across images, 3D scenes and sensor streams. It can be deployed in your own environment.
- NVIDIA Isaac Sim: open-source robotics simulation and synthetic data generation. NVIDIA also publishes free Physical AI Open Datasets on Hugging Face.
- Parallel Domain: deterministic sensor simulation and real-fidelity digital twins (PD Sim and PD Replica) for warehouse robotics.
- Applied Intuition: Dana, an agentic platform for physical AI that covers simulation and evaluation, petabyte-scale ingestion and quality control, and a closed-loop data flywheel. It runs in the cloud, on-premises or air-gapped.
How to choose a robotics data company
- Start from the training stage. Pre-training usually needs volume and variety, so egocentric human video is common (Toloka, Firsthand, Knoxlabs, Labelbox Terra). Post-training and imitation learning need action-labeled demonstrations on your robot or a similar one (Scale AI, Labelbox Terra, Encord, Fax Research). Perception models need accurate 2D/3D labels on your own sensor logs (iMerit, Sama, Segments.ai, Kognic).
- Match the embodiment. Ask whether the vendor can operate your hardware (Scale AI and Labelbox Terra say they can) or will put your gripper or glove on their operators (Knoxlabs). Ask which action and state channels they record, and at what rate.
- Ask for real samples in your format. Several vendors deliver in LeRobot, RLDS, HDF5 or WebDataset; Encord supports MCAP and ROSBAG. Load samples into your own pipeline before signing.
- Check coverage, not just hours. Ask how environments, objects, operators and lighting are spread, and how the vendor finds and fills gaps.
- Demand provenance and consent records. Look for written participant releases, bystander de-identification and clear commercial license terms. Firsthand, Appen and Fax Research document these publicly.
- Run a paid pilot. Most vendors here quote per program. A small scoped pilot is the only reliable way to compare quality and the true cost per usable episode.
- Check security fit. If data must stay in your own cloud, look for VPC or on-premise options: Encord offers VPC, on-premise and air-gapped setups, Foxglove can keep data in your own bucket or run on-premises, and Voxel51 can be deployed in your environment.
FAQ
What is a robotics data company?
A robotics data company collects, labels, verifies, generates or licenses the data robots learn from. That includes teleoperated robot demonstrations, first-person (egocentric) human video, handheld-gripper (UMI) demonstrations, annotated camera, LiDAR and IMU logs, and synthetic data from simulators. This directory groups them into five categories: teleoperation collection, egocentric video, sensor and LiDAR annotation, dataset marketplaces, and simulation/synthetic data.
Which companies provide robot teleoperation data?
Scale AI, Labelbox Terra, Encord and Fax Research publicly offer teleoperation demonstration collection, and Toloka licenses a teleoperation dataset. Knoxlabs offers robot teleop at pilot scale while it expands its Los Angeles facility. NVIDIA publishes free GR00T teleoperation datasets on Hugging Face.
What is the difference between teleoperation, egocentric and UMI data?
Teleoperation data is recorded while a person drives a real robot, so it includes the robot's joint states and actions. Egocentric data is first-person video of people doing tasks, recorded with head-worn and often wrist cameras. UMI data uses a hand-held instrumented gripper, which records gripper pose and aperture without needing a robot.
How much does robotics training data cost?
Most robotics data vendors do not publish prices and quote per program after a pilot. Firsthand bills per validated hour. Toloka says its example HomER egocentric dataset cost $50 on its self-serve platform, but does not publish rates. Segments.ai's Core labeling plan starts at 3,600 hours a year, which it says is equivalent to $2.67 per hour. NVIDIA Isaac Sim is free and open source. See our pricing guide for more.
Can I buy off-the-shelf robotics datasets?
Yes. Toloka licenses off-the-shelf egocentric, UMI, teleoperation and exocentric datasets. Fax Research sources licensed existing material. Appen sells perception sets such as robot-vacuum-perspective images. NVIDIA publishes free Physical AI Open Datasets on Hugging Face.
Can I collect robotics data self-serve without a sales process?
Toloka offers a self-serve egocentric video collection platform with no sales cycle and no minimums. Firsthand and Knoxlabs offer free samples before any commitment, and Segments.ai offers a 14-day free trial of its labeling software. Most other vendors start with a demo or scoping call.
What formats do robotics data vendors deliver in?
Common formats include LeRobot, RLDS, HDF5 and WebDataset for robot-learning episodes, and MCAP or ROSBAG for raw sensor logs. Knoxlabs lists HDF5 and RLDS/LeRobot. Firsthand lists RLDS, LeRobot and WebDataset, plus HDF5, zarr and Rerun .rrd. Encord supports MCAP and ROSBAG. Toloka delivers UMI data as MCAP plus JSON.
Should I use synthetic data instead of real robot data?
Synthetic data from simulators such as NVIDIA Isaac Sim or Parallel Domain is useful for scale, edge cases and testing, but most teams still combine it with real-world demonstrations. The 12 profiled vendors focus on real-world data.
Did you test these companies?
No. This directory is based on each company's public website and documentation. We did not test products, buy data, or verify vendors' volume, quality or customer claims.
Elsewhere in our network: For a straight ranked shortlist, see bestroboticsdatacompanies.com. Related guides: Robotics data providers, training data companies beyond robotics, and bestaidatasets.com for general AI datasets.
Methodology · About · Pricing guide · Get your company featured
Named on this page as a provider? Join the network for a linked profile, Visit buttons and featured placement.