AI data market
Sama physical AI data sourcing questions beyond annotation volume
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Whether you are evaluating Sama or any annotation provider for physical AI work, labeling throughput is only one input. A sourcing review should also establish where the raw recordings came from, who holds rights and consent for them, how sensors were calibrated and synchronized, what the episodes cover including failures and recoveries, where copies are stored during labeling, and which quality figures are provider claims rather than tested results.
Video and sensors
Sama physical AI data
Key takeaways
- Annotation volume measures labeling capacity, not whether the underlying recordings are usable, lawful or representative of the task.
- Sama's public pages, as of October 2026, describe robotics annotation and multi-sensor support; they are the company's own statements and should be tested in a pilot.
- Collection quality turns on calibration, units and frames, time synchronization, episode metadata and coverage of failures and human interventions.
- Rights questions sit with whoever recorded the data: site permission, worker notice and consent, biometric rules and ownership of the recordings.
- Ask where video and point clouds are stored during labeling and for how long, because some workflows copy them into the vendor's storage.
What should a physical AI data sourcing review assess beyond annotation volume?
A physical AI sourcing review should test five things beyond how many frames or points a vendor can label: the origin and rights status of the raw recordings, the consent of the people and sites captured, the technical integrity of the sensor data, the coverage of the task including failures and recoveries, and the custody of the data while it is processed. Annotation volume tells you about throughput. It says nothing about whether the footage was lawfully captured or whether the labels sit on data a robot can learn from.
This guide uses Sama as the worked reference because buyers search for it by name, and it quotes only what Sama publishes about itself, dated and attributed. No affiliation with Sama is implied, and nothing here rates Sama against other providers. The same questions apply to any vendor, in-house team or data licensor.
Which Sama statements are provider claims, and what would substantiate them?
Everything below is what the company says on its own site and documentation as of October 18, 2026. Each statement is a provider claim until a pilot, a contract term or an independent check substantiates it, and the right-hand column suggests how.
| What Sama says | Where it says it | Type of statement | How a buyer could substantiate it |
|---|---|---|---|
| It annotates data for robotics systems, including palletization and depalletization | Robotics and manufacturing page | Service description | Pilot on your own pallet scenes with your label taxonomy |
| Its workflows can handle up to 14 sensors, including cameras, lidar, radar and ultrasonic | Robotics and manufacturing page | Capability claim | Send a multi-sensor sample with your real calibration and sync files |
| A dark line on a pallet can be a gap between boxes or a fold between flaps, and accurate labels on such look-alikes matter | Robotics and manufacturing page | Method statement | Seed the pilot with known ambiguous cases and score them separately |
| A 99% first-batch client acceptance rate, under its SamaAssure program | Data annotation solution page | Performance claim | Ask how acceptance is defined, on what sample, and what the contract commits to |
| A 95% acceptance rate guaranteed in writing | 2D image annotation page | Contractual claim | Read the actual quality clause; the figures differ across pages |
| Video and 3D point cloud assets are processed and stored in Sama's storage regardless of sourcing option | Platform documentation on asset sourcing and hosting | Product documentation | Confirm region, retention and deletion terms in the agreement |
Is buying annotation capacity the same as sourcing the data?
Most annotation providers label data the customer supplies, so the sourcing questions usually land on the buyer, not the labeler. The Sama pages and documentation reviewed for this guide describe annotation and validation of customer assets, including reading them from the customer's own cloud storage; they do not set out where a buyer's raw robot footage should come from. If you need a vendor to collect or source recordings as well, ask that as a separate question with its own terms.
This matters because the two purchases carry different risks. An annotation contract governs labeling quality, turnaround and data handling. A sourcing or licensing arrangement governs whether you may use the recordings at all, for which purposes, and what happens if a person in them objects. A high acceptance rate on labels cannot repair recordings captured without permission or with broken timestamps.
There is one honest exception. If your team already owns rights-cleared footage, collected under its own protocol with calibration files intact, annotation capacity and label quality may genuinely be the main decision. Even then, the custody questions below still apply.
Which collection quality questions should be answered with files?
Ask for evidence in the form of files, not descriptions. Each row names a question, what good evidence looks like, and the open convention it can be checked against.
| Question | Why volume does not answer it | Evidence to request | Reference convention |
|---|---|---|---|
| Are units and coordinate frames declared? | Millions of labeled points are useless if one sensor reports millimeters and another meters | Frame tree, units per channel, handedness | ROS REP 103: SI units, right-handed frames, x forward, y left, z up |
| Are streams time-synchronized? | Labels on a camera frame mean little if the gripper state is from a different instant | Clock source per sensor and measured offsets on a sample | Pilot measurement against the package's own timestamps |
| Is calibration current for each session? | A shifted camera silently moves every 3D label | Calibration files with dates and the sessions they cover | Supplier calibration records |
| Does each episode carry its schema and rate? | Without frame rate and feature definitions, episodes cannot be replayed | Per-dataset schema, frame rate, per-episode lengths and task text | LeRobot dataset v3 meta files: info.json, tasks and episodes |
| Are failures and human takeovers included? | Clean successes alone do not teach recovery | Share of episodes with interventions, how takeovers are marked | LeRobot human-in-the-loop recording of autonomous and human segments |
| Is the collection process documented? | Throughput figures say nothing about how scenes were chosen | Written collection description, collection type, time frame | Croissant RAI fields rai:dataCollection, rai:dataCollectionType |
| Does coverage match deployment? | Volume concentrated in one site or shift overstates diversity | Counts by site, lighting, object class, operator and outcome | Buyer's own task specification |
Who answers the rights, consent and custody questions?
Rights for physical recordings attach to the people, places and contracts involved in the capture, and they need answers before labeling starts. These questions go to whoever made the recordings, whether that is your team, a collection vendor or a business licensing existing footage.
| Question | Who usually holds the answer | What a usable answer looks like |
|---|---|---|
| Who owns the recordings, and under what agreement were they made? | Site operator or collection vendor | A signed agreement naming the owner and permitted AI uses |
| Were the site and equipment owners' permissions obtained? | Facility management, equipment lessors | Written site permission covering recording and reuse |
| Were workers and bystanders notified, and did they consent where required? | Employer, collection vendor | Notice text, consent records and an opt-out path |
| Do recordings capture biometric identifiers such as face geometry? | Collection vendor, privacy counsel | A biometric assessment and, where state law requires, written releases |
| Which personal or sensitive attributes are present? | Data owner | A declared list of sensitive attributes, for example in a Croissant RAI field |
| Are there subcontractors who touch the raw data? | Annotation or collection vendor | A named list of subprocessors and their locations |
Why worker consent questions belong at the top of the list
Footage from factories, warehouses and job sites almost always shows people. In Illinois, the Biometric Information Privacy Act treats a scan of face or hand geometry as a biometric identifier and requires written notice and a written release before a private entity collects one. Other states and countries have their own rules, so the safe sequence is to identify what the recordings contain, then have privacy counsel decide what consent or redaction is needed, before footage goes to any labeling queue.
Redaction is not a shortcut around this. Blurring faces after collection may reduce risk for downstream use, but it does not change whether the original capture was permitted, and some tasks need hands and body pose that cannot be blurred without destroying the signal.
Scoring a palletizing pilot (illustrative): where volume hid the gaps
This scenario is invented to show how the checklist changes a decision; it describes no real vendor, customer or result. A robotics team pilots depalletizing labels on 400 episodes from two warehouses.
On volume, the pilot looks finished: every episode is labeled, and the labels pass the team's spot check on 95% of sampled frames (illustrative). The checklist tells a different story. Calibration files cover only one of the two sites, so 3D labels from the other site cannot be trusted. Only 6% of episodes include a human takeover, all from a single shift. The second site's recordings were made by a contractor whose agreement with the warehouse does not mention reuse for model training. And the video was stored in the labeler's storage for the pilot, which the warehouse's own data policy had not anticipated.
The decision changes from scale up to fix the source. The team asks the second warehouse for calibration and a reuse agreement, commissions additional recovery episodes across shifts, and agrees retention and deletion terms with the labeling vendor before the next batch. None of those steps involves more annotation capacity.
Where the recordings sit while a vendor labels them
Ask exactly which copies exist, where, and for how long, because the answer depends on the asset type and the integration. Sama's documentation describes several hosting options, including reading assets from the customer's own cloud storage, temporary caching in Sama's storage, and direct upload to Sama's storage for the duration of a project. The same documentation states that video and 3D point cloud assets are processed and stored in Sama's storage regardless of the sourcing option.
For physical AI data, which is mostly video and point clouds, that means a copy will usually exist outside your environment while labeling runs. That is common across vendors and not a problem in itself, but it has to match what the data owner agreed to. Put region, retention period, deletion confirmation and subprocessors into the agreement, and check them against the site owner's and workers' terms before the first upload.
When should a safety, privacy or contracts specialist step in?
Bring in a specialist rather than pushing on when the review turns up any of these signals. Recordings include faces, voices or other biometric data and no one can produce the consent records. The footage shows hazardous tasks, near misses or injuries, which raises safety and employment questions beyond data quality. The recordings came through more than one intermediary and the chain of agreements is incomplete. Or the vendor's quality figure cannot be tied to a defined metric and sample.
Narrowing the scope is often the fastest fix. Limiting a first batch to one site with complete calibration and clear agreements produces a smaller dataset you can defend, and the checklist can then be rerun as each new site is added.
Keep a short record of how each question was answered: the file or clause that answered it, who supplied it, and the date. Vendor pages change, figures on them differ from page to page, and staff who ran the pilot move on. A dated evidence log lets a later reviewer see which statements were tested in your own pilot, which rest on a signed term, and which are still only the provider's description of itself, and it turns the next vendor review into an update rather than a fresh start.
What should a request for physical AI recordings specify?
If the gap is the recordings themselves rather than labeling, the request should describe the task, sensors, sites, volume, language for any instructions or narration, delivery format, timing and the licensing terms you need. Recordings are not something SourceX keeps on hand; once a buyer submits those details on the request form, it searches for U.S. businesses whose recordings fit, which puts the rights and collection questions above into the scope from the first conversation.
Frequently asked questions
What belongs in a collection quality and rights checklist for robot data?
Questions on units, frames, calibration, time sync, episode metadata, failure and intervention coverage, and the documented collection process, plus ownership, site permission, worker consent, biometric content, subprocessors and custody. Each answer should point to a file or signed document, not a description.
Does Sama collect physical AI data or only annotate it?
The Sama pages and documentation reviewed for this guide describe annotation and validation of customer-provided assets, including robotics and multi-sensor work. Ask Sama directly whether it offers collection or sourcing for your use case and on what terms.
Is a 99% acceptance rate a measure of data quality?
It is a provider's measure of how often clients accept delivered labels, and Sama's own pages cite different figures in different places. Ask how acceptance is defined, on which sample and under which contract clause, and run your own pilot with known hard cases.
Who is responsible for worker consent in robot training footage?
Usually whoever captured the footage, typically the employer or collection vendor, with the buyer checking the evidence. Where recordings capture face geometry or other biometric identifiers, state laws such as Illinois BIPA can require written notice and a written release before collection.
Where are robot videos stored while an annotation vendor labels them?
It depends on the vendor and the integration. Sama's documentation says video and 3D point cloud assets are processed and stored in its storage regardless of sourcing option, so confirm region, retention and deletion terms in the agreement.
Sources
- Sama says it annotates data for robotics including (de)palletization, handles up to 14 sensors, and cites the box gap versus flap fold example. Source
- Sama's documentation describes hosting options from customer cloud storage to direct upload, and says video and 3D point cloud assets are processed and stored in Sama's storage regardless of sourcing option. Source
- Sama's data annotation page cites a 99% first-batch client acceptance rate. Source
- Sama's 2D image annotation page says it guarantees a 95% acceptance rate in writing. Source
- ROS REP 103 standardizes on SI units and right-handed frames with x forward, y left and z up. Source
- LeRobot dataset v3 stores a canonical schema and frame rate in info.json, task text in tasks, and per-episode lengths and tasks in episode metadata. Source
- LeRobot human-in-the-loop collection records both autonomous and human-controlled segments, including recovery and correction. Source
- Croissant RAI defines rai:dataCollection, rai:dataCollectionType and rai:personalSensitiveInformation fields. Source
- Illinois BIPA treats a scan of hand or face geometry as a biometric identifier and requires written notice and a written release before collection. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.