The physical AI dataset catalog built for buyers, not just researchers.

Egocentric video, teleoperation, RGB-D, point cloud, tactile.

Grasping, manipulation, navigation, household, warehouse, long-horizon.

Franka, UR5, ALOHA, Stretch, Sawyer, xArm, mobile manipulators.

RLDS, LeRobot, HDF5, MCAP, ROS bag, Parquet.

Apache-2.0, MIT, CC-BY, custom, research-only.

Allowed, restricted, unclear, research-only — beyond the license tag.

Side-by-side dataset decisions for buyer-relevant choices.

Ranked nearby datasets per source, scored by overlap on robot, modality, task, and format.

Use Open X-Embodiment for broad cross-robot pretraining; use DROID when real-world manipulation diversity is the deciding factor.

Use Ego4D for broad egocentric activity coverage; use EPIC-KITCHENS for kitchen-specific hand-object action understanding.

Use RoboMimic for imitation-learning trajectory workflows; use Meta-World for multi-task simulated manipulation benchmarks.

Use DROID when in-the-wild manipulation diversity matters; use BridgeData V2 for Berkeley-style behavior-cloning baselines and task generalization references.

Use DROID for broader single-arm real-world manipulation; use ALOHA when bimanual teleoperation and coordinated two-arm tasks are the deciding factor.

Use DROID for real-world manipulation diversity; use RoboMimic for controlled imitation-learning benchmarks and repeatable evaluation workflows.

Use Open X-Embodiment for broad cross-embodiment coverage; use BridgeData V2 when the buyer wants a narrower manipulation baseline with clearer task-family focus.

Use Open X-Embodiment for cross-robot breadth; use ALOHA when bimanual teleoperation and low-cost dual-arm demonstrations are central to the model objective.

Use RLBench for language-conditioned simulated manipulation tasks; use ManiSkill for broader manipulation skill benchmarking and synthetic policy evaluation.

Use RLBench for task-rich simulated manipulation; use RoboSuite for controller experiments and standardized robot-manipulation environments.

Use Meta-World for multi-task reinforcement-learning benchmarks; use CALVIN for long-horizon, language-conditioned manipulation evaluation.

Use CALVIN for long-horizon manipulation in a focused simulated setup; use BEHAVIOR when household task taxonomies and embodied AI planning coverage matter more.

Use Ego4D for broad egocentric activity coverage; use HOI4D when geometry-aware human-object interaction and RGB-D context are more important.

Use EPIC-KITCHENS for kitchen-specific action understanding; use HOI4D when 4D hand-object geometry and broader object interaction are required.

Use DexYCB for dexterous hand-object pose and YCB object grasping; use HOI4D for broader egocentric human-object interaction with RGB-D context.

Use ScanNet for real indoor RGB-D scene reconstruction; use Habitat datasets for simulated embodied navigation and rearrangement tasks.

Use AI2-THOR for interactive household scenes and embodied agent experiments; use BEHAVIOR for broader household activity task specification.

Use ObjectFolder for object-centric geometry, material, and tactile references; use DexYCB for hand-object interaction and pose-labeled grasping examples.

Use RH20T when contact-rich multimodal sensing matters; use DROID when broader in-the-wild manipulation diversity and distributed collection are the deciding factors.

Use AgiBot World to study high-volume robot-data-factory releases; use Open X-Embodiment for cross-institution generalist policy pretraining references.

Use RoboCasa for large-scale kitchen simulation and household scene diversity; use LIBERO for structured lifelong-learning and language-conditioned manipulation benchmarks.

Use RoboSet for multi-view, multi-task kitchen manipulation; use TACO Play when TFDS-compatible Franka kitchen interaction data is the more practical ingestion path.

Use UMI when portable in-the-wild human demonstrations are central; use ALOHA when the buyer needs low-cost bimanual teleoperation data aligned to a robot platform.

Use FurnitureBench for real-world long-horizon assembly demonstrations; use CALVIN for simulated language-conditioned long-horizon manipulation evaluation.

Use RoboTurk as a remote teleoperation collection reference; use RoboSet for a more kitchen-focused real-world multi-task manipulation corpus.

Use LeRobot datasets when the distribution and format ecosystem matter; use Open X-Embodiment when the buyer needs a specific cross-embodiment corpus reference.

Scan the broader robotics dataset surface before narrowing into promoted profiles, comparisons, and custom collection specs.

Track source updates, licensing notes, and buyer-readiness changes that should trigger a renewed review.

Score whether a public source is enough for the model, rights path, modalities, and target environment.

Separate source license language from contributor consent, redistribution, private-space risk, and model-use assumptions.

Turn a public-source gap into a scoped capture request with sample QA, metadata, and delivery requirements.

Compare data providers when the answer is not another public dataset but a better sourcing or capture route.

Use the company index to separate annotation vendors, data engines, marketplaces, and specialist capture teams.

Market context for why physical AI systems need custom, enriched, real-world data beyond generic labeling workflows.

Robotics dataset and tooling context for Hugging Face based collection, sharing, conversion, and training workflows.

A cross-embodiment robotics dataset reference for comparing trajectory scale, robot diversity, and VLA training assumptions.

A large in-the-wild robot manipulation dataset reference for real-world trajectory capture and deployment transfer risk.