ROBOTICS DATA COLLECTION
Robot training data
built around your task.
Real-world robot training data, scoped from your task definition. One partner for egocentric, UMI and teleoperation collection, from a shared brief to a delivered dataset.
01 / WHAT IT IS
Intelligence starts
with experience.
Robotics data collection is the work of producing the demonstrations a robot learning system trains on: recordings of tasks being done, in the form a model can learn from. For physical AI, that means data from the real world rather than from simulation alone, with the object, scene, and behaviour variation that deployment environments contain.
Telepath runs that collection as a service. You define what your model needs to learn. We work backwards from that goal to choose the capture method, the sites, the people, and the variations, and we deliver a dataset structured for your training workflow.
Three capture methods are available and are often combined in one collection: egocentric video for learning from human behaviour, UMI handheld-gripper demonstrations for manipulation policies, and teleoperation data for embodiment-specific demonstrations on your robot.
02 / WHAT WE CAPTURE
What a collection
includes.
Every engagement produces the same core deliverables, whichever capture methods the brief specifies:
- A shared collection brief
- The task definition, capture methods, environments, variations, volume, acceptance criteria, and delivery format, agreed in writing before collection starts.
- Task-aligned demonstrations
- Episodes recorded to the brief, in real work environments, with the variation your model needs built into the collection plan.
- Reviewed data
- Every episode checked against the acceptance criteria, with off-brief captures flagged and thin coverage scheduled for another pass.
- Structured delivery
- Video, pose, action, and metadata streams organised in the structure and formats your team asked for, with a manifest and a coverage report.
- Site access
- Collection through environments across the Middle East and Central Asia: logistics, hospitality, retail, and industrial workplaces.
03 / WHAT IT IS FOR
Who this is for.
Robot learning and physical AI teams who have a model and a task, and are missing the real-world demonstrations to train or evaluate it.
Teams whose lab-collected data does not survive contact with deployment environments and who need object, scene, and behaviour variation they cannot stage themselves.
Teams that want one partner across capture methods, so egocentric, UMI, and teleoperation data on the same task arrive in one consistent structure.
04 / CHOOSING A METHOD
Three methods,
one brief.
The three capture methods record different things and answer different questions. Use the table to decide where to start; the brief can combine methods on the same task.
| Criterion | Egocentric video | UMI demonstrations | Teleoperation data |
|---|---|---|---|
| What is recorded | First-person video from a head- or chest-mounted camera worn by the person doing the task. | Wrist-view video from the handheld gripper, its six-degree-of-freedom pose, and gripper width over time. | The robot’s camera streams and proprioceptive state, paired with the operator’s commanded actions. |
| Who performs the task | A person, with their own hands. | A person, holding the gripper device. | Your robot, driven by a human operator. |
| Robot required | No. | No. The device is configured to match your target gripper and camera placement. | Yes. The collection is scoped around the embodiment you are training. |
| Embodiment match | None. A human reference for how the task is done. | Gripper-level. The data looks like what your robot gripper would record. | Exact. The same observation and action space the robot uses when running a policy. |
| Best suited to | Learning from human video: task understanding, visual representations, hand and object motion. | Manipulation policies from demonstration, and covering object and scene variation cheaply. | Imitation learning on a specific robot, edge-case recovery, and learning from operator intervention. |
| Coverage and speed | No robot involved, so coverage is limited by people and sites rather than platform access. | Faster to collect than teleoperation and covers more variation. | Bound to robot and operator availability; matches the deployment platform precisely. |
| Delivered as | Episodes by task with video, metadata, and variation tags, plus a manifest and coverage report. | Per-episode video, gripper pose, and gripper width, synchronised and segmented, plus a manifest and coverage report. | Observations and actions per episode in your action space, with success and failure labels, and intervention records where scoped. |
05 / WHERE WE COLLECT
Collected across the
Middle East and Central Asia.
Telepath is based in Dubai and runs collections in real workplaces across the United Arab Emirates and the CIS, with sites in Kazakhstan, Uzbekistan and Tajikistan. The data is built for robotics and physical-AI teams in the United States, led by the San Francisco Bay Area.
- Emirates
- Dubai, Sharjah and Ras Al Khaimah.
- Site types
- Laundries, perfume and cosmetics plants, cable factories, furniture factories, plus logistics, hospitality and retail sites.
- Languages
- Briefs and on-site instructions in English, Arabic and Russian.
- Why it matters
- Real objects, real layouts and real interruptions, so the variation your model meets in deployment is already in the training data.
06 / EXAMPLE TASKS
Briefs we can scope
with this method.
Twelve example briefs across four industries. Each shows the kind of task we scope; your own task definition is the starting point for any collection.
LOGISTICS · OBJECT VARIATION / CLUTTER
Order picking
Select different objects from bins and place them into order totes. Vary object geometry, bin layouts, and grasp approaches.
Object selection · Grasping · Placement
Add to my briefLOGISTICS · SORTING / LONG HORIZON
Parcel sorting
Move parcels between containers using visual sorting criteria. Scope different packaging materials, sizes, and work surfaces.
Visual sorting · Transfer · Task sequences
Add to my briefLOGISTICS · BIMANUAL / SEQUENCING
Packing and unpacking
Open containers, arrange items, and handle packaging materials through a complete packing or unpacking sequence.
Container interaction · Bimanual handling · Sequencing
Add to my briefHOSPITALITY · PRECISE PLACEMENT / ARRANGEMENT
Table setting
Arrange plates, cups, and cutlery to a defined layout. Include different table settings and object placements in the brief.
Spatial arrangement · Object orientation · Placement
Add to my briefHOSPITALITY · SOFT OBJECTS / BIMANUAL
Linen folding
Manipulate towels and soft textiles through sequential folds, with variation in material, dimensions, and starting configuration.
Deformable objects · Bimanual actions · Task progression
Add to my briefHOSPITALITY · CLUTTER / TASK SEQUENCES
Surface clearing
Clear cups and objects from a work surface and return them to assigned locations, capturing the order and context of each action.
Scene understanding · Object transfer · Workflow completion
Add to my briefRETAIL · VARIABLE HEIGHTS / PLACEMENT
Shelf replenishment
Transfer products from a tote to shelves across varied heights, layouts, and product shapes.
Reach variation · Object handling · Shelf placement
Add to my briefRETAIL · ORIENTATION / PRECISION
Product facing
Reposition products and align them within a shelf display, with different orientations and starting arrangements.
Object orientation · Fine repositioning · Visual alignment
Add to my briefRETAIL · CATEGORIZATION / OBJECT VARIATION
Product sorting
Group mixed products into bins according to a task brief, including varied packaging and visual selection criteria.
Categorization · Selection · Placement
Add to my briefINDUSTRIAL · SMALL OBJECTS / SEQUENCING
Parts kitting
Select components and assemble a task-specific kit, varying part types, tray positions, and kit requirements.
Component selection · Precise placement · Task sequences
Add to my briefINDUSTRIAL · ORIENTATION / MANIPULATION
Tool handling
Pick up, orient, use, and return tools within an agreed workstation task. Align the capture setup to the actions your team needs.
Tool orientation · Controlled manipulation · Work sequences
Add to my briefINDUSTRIAL · OBJECT VIEWS / HANDLING
Visual parts inspection
Capture how people pick up, rotate, and inspect parts at a workstation, with the inspection criteria defined in the collection brief.
Multi-view observations · Object rotation · Inspection sequences
Add to my brief
EXAMPLE COLLECTION BRIEFS · CUSTOM TASKS WELCOME
07 / HOW A COLLECTION RUNS
Four stages,
one shared brief.
The four stages below run on every collection, whatever the capture method. Each stage has a named output, so both teams know what "done" means before the next one starts.
DefineStart with the capability.
Align on tasks, capture methods, environments, and acceptance criteria. Define what a useful demonstration looks like before collection begins.
YOUR OUTPUT / A SHARED COLLECTION BRIEF
CollectBring the brief into the world.
Coordinate site access, people, and capture setups around the agreed task. Collect the variations and sequences your model needs.
YOUR OUTPUT / TASK-ALIGNED DEMONSTRATIONS
ReviewMake every delivery intentional.
Review captures against the agreed criteria. Identify gaps in the task coverage, flag collection issues, and align on what needs another pass.
YOUR OUTPUT / REVIEWED COLLECTION DATA
DeliverFit your training workflow.
Receive the collection in the dataset structure and format agreed with your team, with the task context and metadata defined in the brief.
YOUR OUTPUT / DATA STRUCTURED FOR YOUR TEAM
FAQ
Robotics data collection, answered.
What is robotics data collection?
Robotics data collection is the production of task demonstrations a robot learning system trains on: video, robot observations, gripper or arm motion, and the metadata around them, recorded in the real world so the resulting dataset reflects the variation a deployed robot will meet.
Which capture methods do you offer?
Egocentric video from a camera worn by the person doing the task, UMI demonstrations from a handheld gripper with pose tracking, and teleoperation data recorded while an operator controls your robot. Many collections combine two of the three on the same task.
How does a project start?
A robotics data collection project starts from your task definition, not from a catalogue. We turn that definition into a written collection brief that fixes capture methods, sites, task variations, episode volume, acceptance criteria, and delivery format. A pilot then runs against the brief before either team commits to a larger collection.
Where do you collect?
Telepath collects robotics data through site access across the Middle East and Central Asia, in real logistics, hospitality, retail, and industrial environments rather than on lab benches. The specific countries and sites for a collection are discussed during scoping, and both can be covered by NDA before any task detail is shared.
Do you sell existing datasets?
No. Telepath does not sell existing or stock datasets. The task cards on this site are example collection briefs, not inventory. Every collection is scoped, collected, reviewed, and delivered for a specific customer against their own task definition, so the data reflects the tasks, environments, and variations that customer’s model actually needs.
How much does robot training data cost per hour?
There is no single per-hour price. Robot training data is quoted per project, and the rate depends on the capture method, the task, how much variation the brief asks for, and the review standard. Every quote states the number of accepted hours or episodes it covers, so the cost per accepted hour is clear before collection starts.
Can I buy off-the-shelf robot training data?
Not from Telepath today. Every collection is scoped around a customer's own task definition and delivered to that customer, because off-the-shelf datasets rarely match a specific robot, gripper and workplace. Pilots are the fastest way to see real data: a small collection on your task, reviewed and delivered within an agreed timeline.
How is pricing structured?
Robotics data collection is priced per project against the collection brief. The main drivers are the capture method, the number of episodes, the task variations required, the number of sites, and the delivery requirements. A pilot is scoped and priced first, so both teams can confirm the setup and review real data before committing to volume.
FROM THE WORLD. INTO YOUR MODEL.