Short answer
Retention, deletion and audit terms decide what happens to a robotics dataset after delivery. Specify how long you may keep raw recordings and derived artifacts, what must be deleted on expiry or termination, how deletion is evidenced, and how the supplier can verify compliance. Match these terms to your training pipeline, because weights, labels and caches complicate deletion. This is not legal advice.
Why post-delivery terms matter more for robotics data
Robotics recordings often include people, workplaces and third-party materials. The data moves into pipelines that create frame extractions, embeddings, labels, tracks and evaluation copies. Each is a copy or derivative that a deletion clause may or may not reach.
Suppliers also have reasons to care. An operating business that licensed footage of its staff typically wants confidence that the footage does not live forever on unknown systems. Buyers want the opposite certainty: that a clause will not force them to retrain or retire a model long after deployment. Writing the clause well is a trade between those positions, and the answer depends on your architecture and the data's sensitivity.
Four obligations to define separately
| Obligation | Question it answers | Common failure |
|---|---|---|
| Retention period | How long may we hold the data? | A fixed term shorter than the model's life cycle |
| Deletion trigger | What starts the deletion clock? | "Termination" undefined for perpetual licenses |
| Deletion scope | Which copies and derivatives? | Silence on caches, backups and extracted frames |
| Verification | How is deletion shown? | A bare statement with no record |
Scope is the part most often left vague. Decide explicitly, for each artifact class, whether it must be deleted, may be kept, or may be kept only in aggregated form.
Illustrative artifact classes
| Artifact | Example treatment a buyer might propose |
|---|---|
| Raw recordings and sensor logs | Delete at term end, or retain for a fixed audit window |
| Extracted frames and clips | Delete with raw recordings |
| Labels and annotations created by the buyer | Retain; strip links back to individual people where feasible |
| Embeddings and features | Negotiate; may allow retention if not reversible to individuals |
| Evaluation copies | Retain under access control for reproducibility, if supplier agrees |
| Model weights | Often proposed for retention; see model weights, derivatives and redistribution |
These are examples, not recommendations. Whether an embedding is "reversible" is a technical question that the contract should not leave to assertion. Ask your engineers what can be reconstructed and write that into a definition.
Deletion and media sanitization
"Deleted" can mean a file-system delete, an overwrite or a physical destruction of media. NIST's media sanitization guidelines, Special Publication 800-88 Revision 2 (published September 2025, superseding Revision 1), are a common reference for describing how storage is sanitized, and a clause can name the method or the standard it must meet (NIST, 2025). Cloud object stores, replicas and backups raise their own questions: a clause may require deletion from active systems within a stated period and from backups on their normal rotation. Ask your infrastructure team which of those you can truthfully certify.
Large datasets add practical points. If recordings arrive on encrypted drives, the clause should say who destroys or returns the drives. For transfer mechanics see delivery and transfer of large datasets.
Where law shapes the clause
This section raises diligence questions, not conclusions.
- Biometric rules. The Illinois Biometric Information Privacy Act defines biometric identifiers to include scans of hand or face geometry, and Section 15(a) requires a publicly available written retention schedule and destruction guidelines, with destruction when the initial purpose is satisfied or within three years of the individual's last interaction, whichever occurs first (Illinois General Assembly). Ask whether any recordings could involve covered identifiers and who holds that schedule. See biometric privacy and workplace video.
- EU data protection. The GDPR treats biometric data processed to uniquely identify a person as a special category and lets Member States set more specific rules for employee data by law or collective agreement (Regulation (EU) 2016/679). If any data subjects or processing sit in the EU, ask counsel how retention limits apply. See EU AI Act and GDPR for robotics data.
Redaction or anonymization does not by itself settle these questions, because they turn on what was captured and under what terms.
Audit rights
Audit rights let the supplier, or a neutral reviewer, check that you follow the license. Expect pushback on scope, because you will not want a supplier inspecting your training infrastructure.
Options, from lightest to heaviest:
- Written compliance certificate at term end.
- Annual self-attestation plus a data-inventory extract.
- Third-party review under confidentiality, limited to data-handling systems.
- On-site audit with notice and cost allocation.
Define frequency, notice period, who pays, what the auditor may see, and how findings are cured. Cross-reference provenance records in provenance and chain of custody, which can make audits cheaper.
Practical drafting notes
Tie the retention period to something both sides can observe, such as a calendar date or termination notice, rather than a loose phrase like "as long as needed". Put the list of artifact classes in a schedule so it can be updated without reopening the contract. Add a notice duty if either party learns of a breach or of a legal hold. State who is responsible for deleting copies held by processors such as cloud providers and annotation vendors, and require that those processors be bound by equivalent terms. Finally, keep a dataset register inside your organization that lists each copy, its location and its owner, because a deletion obligation you cannot locate copies for is one you cannot meet.
Limitations and what these clauses cannot do
Retention and deletion terms do not make data better, and they do not correct a permission problem at the source. They also cannot guarantee that a trained model contains nothing derived from deleted data; whether a clause should require retraining is a commercial choice with real cost. Audit rights verify process; they do not prove that no copy exists elsewhere. And clauses are only as good as the records behind them, so your systems need to be able to locate every copy by dataset ID.
Questions to ask suppliers
- What retention period do you expect, and is it tied to the model life cycle or a calendar term?
- Which artifacts must be deleted on expiry, and which may we keep?
- Do you require deletion from backups, and on what schedule?
- What evidence of deletion do you accept: certificate, log extract, third-party attestation?
- Do the recordings involve individuals whose data is subject to a retention schedule, and who maintains it?
- If a worker or customer withdraws permission, what is the process and the effect on our copy?
- What audit scope are you asking for, and will you accept review of data-handling systems only?
- Who bears audit cost, and how much notice is required?
- If a license is terminated early, how are in-flight training runs treated?
- Are there regulatory holds or insurance obligations that require you to keep records of what we received?
Discuss your requirements
State your retention, deletion and audit expectations when you describe your requirement to SourceX. SourceX reviews the requirement, looks for US businesses that may hold suitable material, checks the data and the supplier's licensing permissions, and coordinates license and delivery. A request does not guarantee a match. Start from the humanoid training data guide or the procurement hub; the licensing terms for robotics training data page covers the wider clause set, and the SourceX trust page covers its trust information.
Discuss your humanoid training-data requirements with SourceX
Sources
- Regulation (EU) 2016/679 (General Data Protection Regulation) — GDPR framing: biometric data as a special category and national employment-data rules, relevant to why retention and deletion terms matter.
- NIST SP 800-88 Rev. 2 Guidelines for Media Sanitization — NIST media sanitization guidance as a reference point for verifiable deletion.
- Biometric Information Privacy Act (740 ILCS 14) — Illinois BIPA requires a public retention schedule and destruction guidelines for biometric identifiers.