Physical AI dataset catalog
Use the catalog to compare source-backed dataset profiles by modality, task, rights signal, consent risk, and deployment fit.
DATASET ALTERNATIVES
If you're using X, what else should you compare?
Pick a dataset and see ranked nearby alternatives by shared modality, robot, task, format, and environment — plus where the alternatives fall short on commercial use, consent, or deployment fit.
DIRECT ANSWER
Alternatives are ranked by overlap with the source dataset, then filtered through buyer-grade signals (commercial use, consent risk, deployment fit) that public benchmarks rarely surface. Open any dataset below to see its top six matches, a side-by-side review table, and a decision guide for when to pick a public alternative versus commission custom data.
36 CURATED PROFILES
Pick the dataset you’re working with
36 of 36 datasets
Open X-Embodiment alternatives
Commercial use unclear · Multi-institution robot demonstration corpus; exact per-task scale varies by contributing dataset.
A large cross-institution collection of robot demonstrations spanning many embodiments and manipulation tasks.
DROID alternatives
Commercial use unclear · Large real-world manipulation corpus; check source for current release counts.
A real-world robot manipulation dataset focused on diverse teleoperated demonstrations outside narrow lab-only settings.
BridgeData V2 alternatives
Commercial use unclear · Robot manipulation demonstrations across multiple tasks; source release describes exact split.
A robot manipulation dataset from Berkeley focused on real-world behavior cloning and task generalization.
RT-1 alternatives
Commercial use unclear · Large language-conditioned robot demonstrations described in the source paper and project materials.
A robotics transformer data release associated with language-conditioned robot manipulation research.
ALOHA alternatives
Commercial use unclear · Task-specific demonstrations released around the ALOHA platform and follow-on projects.
A low-cost bimanual teleoperation platform and dataset family used for imitation learning in dexterous manipulation.
RoboMimic alternatives
Source appears permissive; verify data terms · Benchmark datasets and demonstration formats vary by task suite.
A benchmark and dataset framework for robot imitation learning with standardized tasks and evaluation utilities.
RoboNet alternatives
Commercial use unclear · Multi-robot manipulation dataset; source materials specify exact robot/task counts.
A multi-robot dataset for visual foresight and manipulation policy research.
Meta-World alternatives
Source appears permissive; verify data terms · Benchmark suite of simulated manipulation tasks.
A simulated manipulation benchmark for multi-task and meta-reinforcement learning.
RLBench alternatives
Commercial use unclear · Large simulated task suite; source materials define current task count.
A simulated robot learning benchmark with many manipulation tasks in CoppeliaSim.
CALVIN alternatives
Commercial use unclear · Long-horizon simulated benchmark and demonstrations.
A benchmark for language-conditioned long-horizon robot manipulation in simulated environments.
ManiSkill alternatives
Source appears permissive; verify data terms · Simulation suite with tasks and environments maintained by the ManiSkill project.
A simulation benchmark and toolkit for manipulation skills and embodied AI policy evaluation.
Ego4D alternatives
Commercial use restricted · Large-scale first-person video corpus with annotations; source controls exact access terms.
A large-scale egocentric video dataset focused on first-person human activity understanding.
EPIC-KITCHENS alternatives
Commercial use restricted · Large egocentric kitchen activity dataset with action annotations.
An egocentric video dataset of kitchen activities used for action recognition and human-object interaction research.
HOI4D alternatives
Commercial use unclear · Human-object interaction dataset with 4D annotations; exact release details live on the source site.
A 4D egocentric human-object interaction dataset with RGB-D and pose-oriented annotations.
DexYCB alternatives
Commercial use unclear · Dexterous grasping dataset with YCB objects and pose labels.
A dexterous hand-object interaction dataset centered on grasping YCB objects with 3D annotations.
ScanNet alternatives
Commercial use restricted · Large indoor RGB-D scene reconstruction corpus.
An indoor RGB-D reconstruction dataset used for 3D scene understanding.
Habitat datasets alternatives
Commercial use unclear · Multiple scene and task datasets under the AI Habitat ecosystem.
A family of embodied AI datasets and simulation assets for navigation and rearrangement research.
AI2-THOR alternatives
Source appears permissive; verify data terms · Interactive household simulation scenes and tasks.
An interactive simulated environment for embodied AI agents in household-like scenes.
BEHAVIOR alternatives
Commercial use unclear · Household activity benchmark and simulation assets.
A benchmark for household activities and embodied AI tasks in simulation.
Something-Something V2 alternatives
Commercial use restricted · Large human-object action video dataset; access and current terms are controlled by the dataset host.
A human action video dataset focused on object interactions and temporal reasoning.
Waymo Open Dataset alternatives
Commercial use restricted · Large autonomous driving scenes with cameras and LiDAR.
A large autonomous driving dataset with camera, LiDAR, and labeled traffic scenes.
ObjectFolder alternatives
Commercial use unclear · Object-centric multimodal assets; source materials define current object count and modalities.
A dataset family for object-centric physical properties, geometry, and multimodal perception research.
BC-Z alternatives
Commercial use unclear · Task demonstrations and model references associated with the BC-Z project.
A behavior cloning project focused on zero-shot task generalization for robots.
DexMV alternatives
Commercial use unclear · Dexterous manipulation demonstrations and visual observations; verify source for release details.
A dexterous manipulation dataset focused on multi-view visual observations and hand-object interaction.
Kinetics alternatives
Commercial use restricted · Large action recognition video corpus maintained through public benchmark releases.
A large video action recognition dataset used widely for video model pretraining.
RoboSuite alternatives
Source appears permissive; verify data terms · Simulation tasks and assets for manipulation research.
A simulation framework and benchmark suite for robot manipulation tasks.
RH20T alternatives
Commercial use unclear · Source describes more than 110,000 contact-rich manipulation sequences with visual, force, audio, action, and human demonstration signals.
A real-world contact-rich robot manipulation dataset with multimodal sensing, force, audio, and human demonstration video.
AgiBot World alternatives
Commercial use unclear · Hugging Face organization page describes the Beta release as 1M+ trajectories and 2,976.4 hours across 217 tasks, 87 skills, 3,000+ objects, and 100+ real-world scenarios.
A large-scale real-world robot manipulation dataset family for fine-grained manipulation, tool use, and multi-robot collaboration.
RoboCasa alternatives
Commercial use unclear · RoboCasa365 source materials describe 365 everyday tasks, 2,500 kitchen environments, 600+ hours of human demonstration data, and 1,600+ hours of synthetic demonstrations.
A large-scale kitchen simulation framework and dataset family for everyday manipulation tasks in diverse household environments.
LIBERO alternatives
Commercial use unclear · Benchmark datasets are organized around multiple LIBERO task suites, including spatial, object, goal, and long-horizon manipulation variants.
A benchmark suite for lifelong robot learning and language-conditioned manipulation tasks.
RoboSet alternatives
Commercial use unclear · Source describes 30,050 trajectories, including 9,500 collected through teleoperation, across 12 skills and 38 tasks with four camera views.
A real-world multi-task kitchen manipulation dataset with teleoperated and kinesthetic demonstrations.
RoboTurk alternatives
Commercial use unclear · Project materials describe over 100 hours of real robot data and thousands of successful manipulation demonstrations collected through remote users.
A large-scale teleoperation data collection platform and dataset family for robot manipulation tasks.
UMI alternatives
Commercial use unclear · Project materials emphasize portable in-the-wild data collection and fast demonstrations for tasks such as cup manipulation, dish washing, cloth folding, and dynamic tossing.
Universal Manipulation Interface is an in-the-wild human demonstration framework for transferring portable gripper data to robot policies.
FurnitureBench alternatives
Commercial use unclear · Documentation describes 219.6 hours and 5,100 successful furniture assembly demonstrations collected with controller and keyboard inputs.
A real-world long-horizon furniture assembly benchmark with successful demonstration data.
TACO Play alternatives
Commercial use unclear · TensorFlow Datasets documentation lists TACO Play as Franka kitchen interaction data with train and test splits and a 47.77 GiB dataset size.
A kitchen robot manipulation dataset with Franka arm interaction data available through TensorFlow Datasets.
LeRobot datasets alternatives
Commercial use unclear · LeRobot documentation describes a standardized dataset ecosystem on Hugging Face Hub using Parquet for tabular data and MP4 for video observations.
A Hugging Face robotics dataset ecosystem and standardized dataset format for multimodal robot learning data.
RESEARCH PATHS
Use this record as part of a broader dataset review
A dataset record is only useful when it connects into the rest of the buyer workflow. The next review step is usually not another summary; it is a fit check, rights triage, source comparison, or custom bounty spec that names the missing proof.
For physical AI teams, the hard question is whether the public source can support a specific model objective under real deployment constraints. That requires adjacent dataset records, tools, comparisons, and sourcing paths, plus external references that a reviewer can open and challenge.
Use the links below to keep the review grounded. Start broad when discovery is incomplete, move into profile and comparison pages when the candidate source is known, and switch to custom collection when the blocker is rights, consent, geography, robot embodiment, or target environment coverage.
INTERNAL LINKS
Continue the buyer workflow
Hugging Face robotics index
Scan the broader robotics dataset surface before narrowing into promoted profiles, comparisons, and custom collection specs.
Dataset changelog
Track source updates, licensing notes, and buyer-readiness changes that should trigger a renewed review.
Dataset fit checker
Score whether a public source is enough for the model, rights path, modalities, and target environment.
License risk checker
Separate source license language from contributor consent, redistribution, private-space risk, and model-use assumptions.
Data spec generator
Turn a public-source gap into a scoped capture request with sample QA, metadata, and delivery requirements.
Vendor alternatives hub
Compare data providers when the answer is not another public dataset but a better sourcing or capture route.
Data annotation companies
Use the company index to separate annotation vendors, data engines, marketplaces, and specialist capture teams.
EXTERNAL REFERENCES
Source context to verify
Scale AI physical AI data engine
Market context for why physical AI systems need custom, enriched, real-world data beyond generic labeling workflows.
LeRobot documentation
Robotics dataset and tooling context for Hugging Face based collection, sharing, conversion, and training workflows.
Open X-Embodiment
A cross-embodiment robotics dataset reference for comparing trajectory scale, robot diversity, and VLA training assumptions.
DROID dataset
A large in-the-wild robot manipulation dataset reference for real-world trajectory capture and deployment transfer risk.
TRUELABEL ROUTING
None of the alternatives fit?
If the public catalog can't close the gap on rights, geography, embodiment, or task coverage, commission custom data with sample QA and acceptance gates.