Explore the data. Know its source.
Human demonstrations and robot-native trajectories from the research community. Each collection belongs to its publisher; the published scale below is separate from CosmicBrain’s data inventory.
Human data, built for robot learning.
EgoVerse connects human demonstrations with processing, training and evaluation tools. Its explorer lets you inspect episodes, tasks, camera views and metadata across contributing sources.
Open the EgoVerse explorer- 01Egocentric videoSee the human’s view
- 02Poses & task contextInspect recorded signals
- 03Human-to-robot learningEvaluate transfer on robots
8 collections · Public source references reviewed October 2, 2026
EgoVerse
EgoVerse consortium
Human demonstrations from homes, workshops and labs, curated for learning how human behavior transfers to robots.
- Published scale
- Living releaseContinuously expanded by the consortium. Project and explorer snapshots report different totals.
- Task examples
- Object placement · Grocery packing · Household manipulation
- Egocentric video
- Camera poses
- Head tracking
- Language annotations
Episode-specific licenses
Check each episode's license before reuse. The repository's MIT code license does not grant blanket dataset rights.
Publisher’s usage termsEgoSuite / EgoDemo
Lightwheel
A 50-hour sample of EgoSuite human activity, spanning its annotated subsets and two raw-video variants.
- Published scale
- 50-hour sampleEgoSuite's launch report lists 10,000 hours released and 100,000 hours planned; EgoDemo is the sample.
- Task examples
- Cooking · Tool use · Packing · Human-to-robot transfer
- Head-view video
- Wrist video (subset)
- Hand pose
- Body pose (subset)
- Semantic annotations
commercial-training-no-resale-v1.0
Custom terms support commercial training and restrict resale. Training rights do not establish permission to republish previews.
Publisher’s usage termsHoloAssist
HoloAssist research team
Collaborative physical tasks with an instructor verbally guiding a performer wearing a mixed-reality headset.
- Published scale
- 169 hoursThe project page reports 169 hours; the ICCV 2023 paper reports 166 hours.
- Task examples
- Mistake detection · Intervention prediction · Hand forecasting
- RGB
- Depth
- Hand pose
- Head pose
- Eye gaze
- IMU
- Audio
- Action / conversation annotations
CDLA–Permissive–2.0
The publisher releases the dataset under CDLA–Permissive–2.0. Include the agreement text when sharing data.
Publisher’s usage termsCaptainCook4D
CaptainCook4D research team
Kitchen recipe recordings that capture correct procedures and deliberate errors, with step and action annotations.
- Published scale
- 384 recordings94.5 hours, with 5.3K step annotations and 10K fine-grained action annotations.
- Task examples
- Procedural error recognition · Step localization · Procedure learning
- Egocentric video
- Step annotations
- Action annotations
- Error labels
Apache 2.0 · dataset
The publisher explicitly licenses the dataset under Apache 2.0 and documents participant consent.
Publisher’s usage termsDROID
DROID Dataset Team
Franka Panda manipulation demonstrations collected across varied real-world scenes using a shared robot platform.
- Published scale
- 350 hours76K demonstration trajectories across 564 scenes and 86 tasks, as reported by the project.
- Task examples
- Object manipulation · Policy learning · Scene generalization
- Exterior / wrist cameras
- Depth
- Camera calibration
- Robot actions / state
- Language instructions
CC BY 4.0 · data
The official paper releases the full dataset under CC BY 4.0. Attribute the dataset and indicate changes when sharing.
Publisher’s usage termsBridgeData V2
BridgeData V2 research team
WidowX manipulation trajectories for learning tasks conditioned on natural-language instructions or goal images.
- Published scale
- 60,096 trajectories50,365 teleoperated demonstrations plus 9,731 scripted rollouts, collected in 24 environments.
- Task examples
- Pick and place · Doors and drawers · Cloth folding · Goal-conditioned learning
- RGB
- Depth (subset)
- Robot actions / state
- Language instructions
CC BY 4.0 · data
The publisher provides all data under CC BY 4.0. Attribute the dataset and indicate changes when sharing.
Publisher’s usage terms
AIST Bimanual Manipulation
AIST · Motoda and collaborators
Dual-arm leader/follower demonstrations on an ALOHA platform, with synchronized observations and joint control signals.
- Published scale
- 10,000+ episodes100+ tasks in the v1 release, as reported by the project.
- Task examples
- Cable insertion · Assembly · Bimanual manipulation
- 4-camera RGB
- Depth
- 14-DoF joints / actions
- Language prompts
CC BY 4.0 · project work
The official project licenses its work and previews under CC BY 4.0. Retain credit and license information; check the downloaded release's accompanying terms.
Publisher’s usage termsHO-Cap
UT Dallas & NVIDIA · Wang and collaborators
Human hand-object interactions captured with egocentric HoloLens and multiple RGB-D views, with 3D hand and object pose annotations.
- Published scale
- Multiview human captureEgo and external camera views. See the publisher's release for sequence and split details.
- Task examples
- Pick and place · Handovers · Object use
- Egocentric RGB
- Multiview RGB-D
- 3D hand pose
- 3D object pose
CC BY 4.0 · data
The publisher releases the dataset under CC BY 4.0. The toolkit's separate software license does not change the dataset license.
Publisher’s usage termsEach collection includes a real data sample or the publisher’s official project preview, credited at the point of use. Preview access is separate from dataset access. Available subsets and usage terms are set by each publisher.
Experience from operations. Scoped for your model.
We collect real deployment data and have hundreds of thousands of hours of teleoperation data available for model teams. Request the inventory to review coverage, sample availability, formats and licensing.
Robot teleoperation
Human-guided robot experience. Discuss target tasks, embodiments, observation and action signals, operator involvement and annotation needs.
For imitation learning & model researchLive deployment data
Experience from robots working at actual sites. Discuss task context, operating conditions, interventions and the evidence required for your evaluation.
For adaptation & field evaluationCoverage and usage rights are confirmed per collection and engagement.
The view. The action. The context.
Choose data by the supervision your model needs. A first-person video, a tracked human hand and a commanded robot action are different signals.
Egocentric demonstrations
First-person task execution, with language, hand poses or other sensors where provided. Useful for studying activity, objects and human manipulation. Human poses require an explicit mapping to a robot’s action space.
Teleoperated trajectories
Robot observations paired with control signals and state. Check coordinate frames, units, control frequency, gripper conventions, calibration and synchronization before combining embodiments.
Deployment experience
Tasks in operating environments, with control ownership, intervention logs and outcomes. Separate demonstrations from autonomous trials, and keep sessions and sites disjoint when evaluating generalization.