Short answer
There are three ways to source robot training data: license existing business assets, commission new collection, or engage specialist partners for robot-specific data. Existing assets are quickest to evaluate but were recorded for other purposes, while commissioned collection fits your specification but needs sites, people and time. Specialist capture supplies robot actions, motion capture, tactile or force data on a defined embodiment. Most humanoid programs combine all three.
The decision this page settles
A data-program lead usually faces the route question before anyone has written a specification: should we look for something that already exists, pay to record it, or go straight to robot data? The answer differs by layer of the stack, so the useful question is which route serves each requirement. This page compares the three routes across the whole section; for the narrower choice between licensing workplace video and recording new video, see existing footage vs commissioned recording.
The three routes compared
| Criterion | Existing business assets | Commissioned collection | Specialist robot-data partners |
|---|---|---|---|
| What you get | Video, documents, images and records made during normal operations | New human recordings designed for training | Robot observations and actions, mocap, tactile or force data |
| Control over viewpoint, sensors and labels | None at capture; labels added later | High | High, tied to one embodiment or rig |
| Robot actions | Only from businesses already running robots, usually under vendor terms | Only if a handheld device or robot is used | Yes |
| Rights picture | Inherited; may involve vendors, clients or equipment makers | Designed in, including consent | Usually simple work-for-hire, plus performer releases for mocap |
| What drives cost | Search, rights review, preparation, annotation | Sites, demonstrators, equipment, calibration, QA | Hardware, operators, resets, episode QA |
| What drives lead time | Supplier approval and rights checks | Site access, recruitment, pilot | Hardware availability and protocol setup |
| Typical weakness | Coverage and viewpoint fixed by the original purpose | Narrower than real operations; staged risk | Narrow task and scene variety per hour of effort |
No route is cheapest in general. An archive that needs heavy redaction and relabeling can cost more per usable hour than a well-run pilot recording, and robot data that cannot transfer to your embodiment is expensive whatever it cost to collect.
Route 1: existing business assets
Choose this route when the layer you need is perception, task understanding, language, objects, failures or evaluation, and when real-world variety matters more than a specific camera position. Operating businesses already hold training libraries, process recordings, SOPs, product catalogs, inspection archives and system logs. These assets show real sites, real pace and real exceptions in volume.
Their limits come from their origin. Viewpoints were chosen for training staff or monitoring a line; labels, if any, serve a business workflow; consent and contracts were written for another use. Expect to add step labels, check rights document by document, and accept gaps you cannot fill. Do not expect robot actions or forces from this route, except robot logs at sites that already run robots.
SourceX looks for US businesses that hold the data a buyer describes, checks the data and the supplier's licensing permissions, and agrees pricing and allowed uses in a license. Nothing is contracted until a supplier agrees.
Route 2: commissioned collection
Choose this route when you need a specific viewpoint (head, wrist, multi-view), extra sensors such as depth or IMUs, tracked hands, a task script, deliberate failures, or consent written for AI training. It also suits evaluation sets that must be held out by site or operator from the start.
Design takes real effort. Ego4D, a purpose-built egocentric corpus, recruited consenting participants, applied de-identification and reports more than 250,000 hours of annotator effort (Grauman et al., 2021). DROID shows the same discipline on the robot side: 50 data collectors across North America, Asia and Europe worked over 12 months with a replicated rig (Khazatsky et al., 2024). Your program will be smaller, but the same elements apply (protocol, calibration, consent, quality control and a pilot); see scoping a pilot collection.
The site describes a separate physical-world path for new recordings of hands-on work coordinated with businesses, where each opportunity has its own requirements. If no existing asset fits, SourceX can scope a commissioned collection of new recordings.
Route 3: specialist robot-data partners
Choose this route for anything a control policy learns directly: robot joint states and actions, end-effector trajectories, contact forces, tactile signals, whole-body motion for retargeting. These come from robots, teleoperation rigs, handheld capture devices or instrumented setups, and they require a specification of embodiment, action space, control frequency, calibration and synchronization; start with teleoperation data collection requirements.
Protocols vary in cost and coverage. Mobile ALOHA used whole-body teleoperation with 50 demonstrations per task and reports that co-training with existing static ALOHA data raised success by up to 90% (Fu et al., 2024). Handheld grippers trade robot presence for portability: the Universal Manipulation Interface collects demonstrations without a robot and reports hardware-agnostic policies deployable across platforms (Chi et al., 2024). SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also arrange annotation and engage specialist capture partners, scoped case by case.
Before commissioning robot data, check what exists. Open X-Embodiment pooled more than one million trajectories from 22 embodiments and reports positive transfer across platforms (Open X-Embodiment Collaboration, 2023); cross-embodiment robot datasets explains when that helps a humanoid. Data generation can also stretch a seed set: MimicGen generated more than 50,000 demonstrations from about 200 human ones (Mandlekar et al., 2023), though the seed still had to be collected.
How the routes combine
Most programs use each route for a different layer, and the routes change each other's value. EgoMimic reports that, in its setting, an extra hour of tracked human hand data was worth more than an extra hour of robot data when the two were co-trained (Kareer et al., 2024). That does not make robot data optional; it changes how much you buy. The boundary between the two kinds of demonstration is covered in human demonstrations vs teleoperated robot demonstrations.
Illustrative route matrix
| Requirement (synthetic program) | First route to try | Fallback | Why |
|---|---|---|---|
| Visual pretraining on 10 workplace task families | Existing assets | Commissioned capture at 3 sites | Variety matters more than viewpoint |
| Step labels and instructions for the same tasks | Existing SOPs plus annotation | Commissioned narration | Procedures already name the steps |
| Hand-tracked demonstrations for co-training | Commissioned collection | None from archives | Needs calibrated cameras and 3D hands |
| Policy data on the target humanoid | Specialist teleoperation | Cross-embodiment data first | Only robot data carries the actions |
| Fastening with torque feedback | Specialist force/torque capture | Instrumented tools | Video has no force signal |
| Held-out evaluation at unseen sites | Commissioned collection | Existing assets with enforced split | The split must be designed, not inferred |
Decision rules
- Start with existing assets when the need is perception, task structure, language or evaluation, and rights can be confirmed.
- Commission when you need a viewpoint, sensor, label or consent that no archive will have, or a clean holdout.
- Go specialist when the model must learn actions, forces or touch on your embodiment.
- Pilot every new route on a small scope before committing the budget, and write the acceptance tests first.
Limits no route removes
Rights still need checking on every route: commissioned work needs consent and site permission, and specialist work needs releases and clear ownership of derived data. Evaluation data must be separated by site, operator, object or time whichever route supplies it. And no route has a universal volume target; requirements depend on architecture, task, embodiment and intended use. Write those down in the requirements specification template before asking anyone for supply.
Questions to ask before choosing a route
- Which layer of our stack does this requirement feed, and does it need robot actions or forces?
- Does any existing asset show the task at our target sites or comparable ones, and from which viewpoint?
- What would we have to add to an existing asset (labels, calibration, consent) and is that possible after the fact?
- If we commission, which sites can host recording, and what disruption and safety review will they require?
- If we go specialist, which embodiment, action space and control rate must the data match?
- Can public or cross-embodiment robot data cover part of the need, and do its licenses allow commercial use?
- Who will own annotations, pose estimates and retargeted trajectories derived from each route?
- Which route supplies the evaluation set, and how will it stay out of training?
Discuss your requirements
You do not need to have chosen a route before you ask. Describe the requirement to SourceX, layer by layer, with the uses you need licensed. SourceX reviews it, investigates whether US businesses hold suitable material, and follows up about potential sources; a request does not guarantee a match. The procurement hub and the humanoid training data guide map the rest of the decisions.
Explore sourcing and licensing options with SourceX
Sources
- Open X-Embodiment: Robotic Learning Datasets and RT-X Models — pooled robot data and cross-platform transfer.
- Ego4D: Around the World in 3,000 Hours of Egocentric Video — effort and consent design in purpose-built collection.
- DROID: A Large-Scale In-The-Wild Robot Manipulation Dataset — scale of a coordinated robot collection effort.
- Mobile ALOHA: Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation — teleoperation protocol and co-training result.
- Universal Manipulation Interface: In-The-Wild Robot Teaching Without In-The-Wild Robots — handheld demonstration capture.
- MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations — generated demonstrations from a seed set.
- EgoMimic: Scaling Imitation Learning via Egocentric Video — relative value of human and robot data when co-trained.