SCENARIO SAMPLE
Household & hospitality
Restocking, room service, linen and object handling
Request a sample review01 / OBSERVE. ACT. LEARN.
Real-world data. Purpose-built for learning.
Data collection, annotation and dataset sourcing around the tasks your models need to understand.
THE RIGHT STARTING POINT
Start with the task or the gap in your dataset. Select a collection setup, define the processing pipeline, or request existing data with a sample and clear usage rights. Observations, actions, environments and acceptance criteria are scoped together.
BROWSE THE DATA PRODUCT
Start with a sample you can inspect.
Review the environment, action boundaries, metadata and delivery shape before committing to a batch.
SCENARIO SAMPLE
Manipulation & tools
Pick, place, sort and use tools across changing scenes
Request a sample review
SCENARIO SAMPLE
Robot interaction
Platform-specific observations, actions and outcomes
Request a sample reviewANNOTATION LAYERS
From an episode to training-ready structure.
The same recording can support different levels of supervision. Select the layers that match your model, evaluation plan and budget.
Configure annotation01
Episode
Stable IDs, recording setup, scene and task context
02
Action timeline
Task, step and atomic action boundaries with timestamps
03
Semantic layer
Objects, locations, goals, outcomes and language descriptions
04
Delivery layer
JSON, Parquet, RLDS, LeRobot or a customer-defined schema
DELIVERY & ACCEPTANCE
Make coverage and limitations visible.
Every project defines the accepted unit, task and scene coverage, required modalities, exclusions, quality rules and usage rights. Raw recording duration is reported separately from accepted training data.
Sample manifestAnnotation guideQuality reportChecksums & versionUsage rightsKnown limitations
Browse data offerings WHAT WE CAN SCOPE
Details that make the difference.
01 /
Human egocentric data
Robot-free first-person recordings capture the diversity of human work. Stereo and multi-camera configurations support task, step and action annotation.
02 /
Robot-embodied data
Platform-specific demonstrations connect camera observations with available robot states, actions and outcomes. Exact channels are validated on a sample.
03 /
Custom task distributions
Scope variations in objects, placement, lighting and task sequence. Define which failures and recovery attempts belong in the dataset.
04 /
Real service environments
Start with concrete workflows in hospitality and sports operations: restocking, linen handling, equipment organization and tool interactions.
05 /
Licensing that follows the use case
Agree training and evaluation use, redistribution, derivatives, exclusivity and geographic handling before an order. Availability is confirmed per collection.
06 /
A sample before a commitment
Review representative episodes, labels, recording configuration and a field dictionary before setting batch volume and delivery milestones.
DEFINE THE DELIVERABLEA clear handover.
A clear handover.
No missing context.
- Versioned episodes and stable identifiers
- Data dictionary, units and recording configuration
- Agreed labels, task instructions and outcome definitions
- Quality report, exclusions and acceptance results
- Checksums, delivery manifest and release notes
- Documented usage rights and retention arrangements
FIND YOUR STARTING POINT
Full catalogExplore the offering.
Human egocentric data
First-person demonstrations of real work, organized into tasks, steps and actions. Build a richer understanding of how people interact with objects and environments.
Robot manipulation data
On-robot task episodes with platform configuration and available observations, states and actions. Align the recording schema with the policy you are training.
LET'S GET SPECIFICYour questions,
Your questions,
answered.
What is robot-free data?
Human egocentric data recorded without requiring a robot during collection. It can describe human actions and intentions, but robot control signals and retargeting are separate requirements.
Can human and robot data be combined?
Yes, a project can cover related tasks in both modalities. We agree correspondence, schema and evaluation criteria; we do not assume that human motion transfers directly to any robot.
Do you sell existing datasets?
Request a current sample list. Existing availability, license terms and usable volume are verified for the requested tasks; new collection can fill gaps.
How do you measure volume?
By the accepted unit stated in the order, such as usable recording hours or validated episodes. Raw recording duration is reported separately from accepted training data.
YOUR NEXT MOVE
Let's build something