llms.txt
Compare with other sites' llms.txt files →# Claru
> Claru provides purpose-built training data for physical AI, robotics, and embodied intelligence — 4M+ human annotations, 386K+ egocentric video clips, 34K+ contributors across 70+ countries.
## Core Pages
- [Training Data for Physical AI](https://claru.ai/): Homepage. Claru is the only training data company 100% focused on physical AI. 4M+ annotations, 386K+ egocentric video clips, 34K+ contributors in 70+ countries.
- [Data Catalog](https://claru.ai/data-catalog): Browse 100+ datasets spanning egocentric video, teleoperation recordings, game environments, and multi-view captures with annotation layers.
- [Solutions](https://claru.ai/solutions): Overview of data solutions for robotics, embodied AI, VLA models, and world models — from raw capture to enriched, annotated datasets.
- [About Claru](https://claru.ai/about): Company background, team, mission. Founded 2024, focused on closing the data gap for frontier robotics labs.
- [Compare Data Providers](https://claru.ai/compare): Side-by-side comparisons of Claru vs Scale AI, Appen, Labelbox, and 90+ other data providers for robotics use cases.
## Training Data Solutions
- [Egocentric Video Data](https://claru.ai/solutions/egocentric-video-data): 386K+ first-person video clips from real-world activities. Kitchen, workshop, warehouse, outdoor, and office environments.
- [Manipulation Trajectory Data](https://claru.ai/solutions/manipulation-trajectory-data): Teleoperation recordings with 6/7-DoF action labels for training robot manipulation policies.
- [Teleoperation Data](https://claru.ai/solutions/teleoperation-data): Remote-operated robot demonstrations with synchronized camera + action streams for VLA and behavioral cloning training.
- [RLHF Labeling](https://claru.ai/solutions/rlhf-labeling-data): Human preference annotations for reinforcement learning from human feedback on robot behaviors.
- [Video Annotation](https://claru.ai/solutions/video-annotation-data): Frame-level and temporal annotations for video understanding — bounding boxes, segmentation masks, action labels.
- [Red Teaming](https://claru.ai/solutions/red-teaming-data): Adversarial evaluation and safety testing for AI models with human evaluators.
## Robotics Knowledge Base
- [Glossary](https://claru.ai/glossary): 70+ technical terms defined in depth — VLA, behavioral cloning, action space, diffusion policy, sim-to-real, and more.
- [How-To Guides](https://claru.ai/guides): 40 step-by-step guides for robotics data — collecting teleoperation data, training diffusion policies, building datasets.
- [Model Pages](https://claru.ai/models): Training data requirements for 25 robot learning models — OpenVLA, RT-2, Octo, pi-zero, ACT/ALOHA, GR00T N1, and more.
- [Task Pages](https://claru.ai/training-data): Data requirements for 30 robotics tasks — grasping, assembly, navigation, dexterous manipulation, and more.
- [Dataset Pages](https://claru.ai/datasets): 25 dataset profiles with technical specs, modalities, annotation layers, and comparison tables.
- [Benchmark Pages](https://claru.ai/benchmarks): 15 robotics benchmarks — RLBench, CALVIN, ManiSkill, MetaWorld — with sim-to-real data bridging guides.
- [Format Guides](https://claru.ai/formats): 15 robotics data formats — RLDS, HDF5, LeRobot, zarr, ROS bag, MCAP — with specs and conversion guides.
- [Industry Pages](https://claru.ai/industries): Training data needs by vertical — healthcare, warehouse, defense, automotive, manufacturing, agriculture, and more.
## Blog
- [VLA Training Data Volume Guide](https://claru.ai/blog/vla-training-data-volume): How much demonstration data VLA models need — numbers from RT-2, OpenVLA, Octo, and pi-zero papers.
- [VLM vs VLA Explained](https://claru.ai/blog/vlm-vs-vla): Technical comparison of vision-language models and vision-language-action models for robotics.
- [Sim-to-Real Gap](https://claru.ai/blog/sim-to-real-gap): Why simulation data alone cannot replace real-world demonstrations for robot manipulation policies.
- [Physical AI Stack](https://claru.ai/blog/physical-ai-stack): The full technology stack for building physical AI systems — data, models, simulation, deployment.
- [Data Enrichment Pipeline](https://claru.ai/blog/data-enrichment-pipeline-physical-ai): How Claru enriches raw video with depth maps, pose estimation, segmentation, and optical flow.
- [VLA Training Data Guide](https://claru.ai/vla-training-data-guide): Comprehensive guide to building datasets for vision-language-action model training.
## Academic Dataset Alternatives
- [Ego4D Alternative](https://claru.ai/compare/ego4d-alternative): Custom-collected, robot-ready data compared with the Ego4D consortium's dataset, which permits specified commercial uses.
- [Open X-Embodiment Alternative](https://claru.ai/compare/open-x-embodiment-alternative): Commercial data for cross-embodiment robot learning beyond the OXE academic dataset.
- [EPIC-KITCHENS Alternative](https://claru.ai/compare/epic-kitchens-alternative): Kitchen activity data with commercial licensing and broader task coverage.
- [RLBench Alternative](https://claru.ai/compare/rlbench-alternative): Real-world manipulation data to bridge RLBench simulation policies to physical robots.
## Case Studies
- [Egocentric Video Data Collection](https://claru.ai/case-studies/egocentric-video-collection): 386K+ first-person clips across three parallel pipelines — GoPro wearable, smartphone, and activity-specific — with same-day QA and weekly delivery.
- [Workplace Egocentric Data](https://claru.ai/case-studies/workplace-egocentric-data): 9.8K+ accepted hours of first-person video from inside active workplaces — 17 categories, 42 countries — for general-purpose robotics.
- [Household Egocentric Video for Humanoid Robots](https://claru.ai/case-studies/household-egocentric-video-humanoid-robots): 344K+ accepted hours of household egocentric video at humanoid camera height — 220+ tasks across 32 countries, commercially licensed.
- [Narrated Egocentric Video](https://claru.ai/case-studies/narrated-egocentric-video): 4,700+ hours of egocentric video with synchronized spoken narration across 26 countries for language-conditioned learning.
- [Video SFT Data Curation](https://claru.ai/case-studies/video-sft-data-curation): 286,712 curated video SFT labels for a frontier video-generation lab — 276,883 delivered in the final 3 days across 233 annotators.
- [Robot Action Caption Verification](https://claru.ai/case-studies/robot-action-caption-verification): 47,251 two-pass verifications of robot action captions on egocentric video; the redesigned binary flow delivered 44,555 in 12 days.
- [Video Quality Annotation at Scale](https://claru.ai/case-studies/video-quality-at-scale): 976K+ human quality assessments for RLHF reward models and model selection.
- [Video Content Classification](https://claru.ai/case-studies/video-content-classification): 105,000 video clips classified in 7 days with zero rework after a mid-project framework redesign.
- [Object Identity Persistence](https://claru.ai/case-studies/object-identity-persistence): 1.07M+ cross-segment identity verifications teaching video models to maintain object identity.
- [Game-Based Data Capture](https://claru.ai/case-studies/game-based-data-capture): 10,000+ hours of synchronized gameplay footage paired with timestamped control data from 31 indexed titles.
- [Video Model Evaluation](https://claru.ai/case-studies/video-model-evaluation): 39,000 pairwise human evaluations across 51 video models using ELO ranking.
- [Prompt Enhancement Benchmark](https://claru.ai/case-studies/prompt-enhancement-benchmark): 180 prompts assessed across text and video with human evaluation and statistical significance testing.
- [Generative AI Safety](https://claru.ai/case-studies/generative-ai-safety): 241,000+ safety annotations measuring residual risk in automated moderation.
- [Red-Teaming Moderation](https://claru.ai/case-studies/red-teaming-moderation): Building and stress-testing a production content moderation system with sub-2% rejection.
- [Fashion AI Annotation](https://claru.ai/case-studies/fashion-ai-annotation): 1,000+ annotators labeling 3 million fashion images with multi-layer QA.
- [Bellwood Infrastructure Monitoring](https://claru.ai/case-studies/bellwood-infrastructure-monitoring): 75+ hours of geo-tagged dashcam footage enriched with automated defect detection across 10 infrastructure categories.
## Key Facts
- Founded: 2024
- Legal entity: Reka AI Inc. (d/b/a Claru AI)
- Focus: 100% physical AI and robotics training data (not general-purpose labeling)
- Data scale: 4M+ human annotations, 386K+ egocentric video clips, 100+ datasets
- Workforce: 34,000+ contributors across 70+ countries
- Partnerships: 5+ frontier AI lab partnerships
- Key datasets: Egocentric video (kitchen, warehouse, workshop, outdoor), teleoperation recordings, game environments, multi-view manipulation
- Enrichment: Every clip ships with depth maps (Depth Anything V2), pose estimation (ViTPose), segmentation (SAM3), optical flow (RAFT), action labels
- Delivery formats: RLDS, HDF5, zarr, LeRobot, WebDataset, ROS bag, MCAP
- Contact: contact@claru.ai
## Contact
- Website: https://claru.ai
- Email: contact@claru.ai
- Book a call: https://calendly.com/claru