Short answer
A robotics data license should treat derived annotations, poses and trajectories, simulation assets, synthetic data, model weights and the raw data as separate outputs, each with its own answer on ownership, permitted use, redistribution and survival after the term ends. Name every output you expect to create and who may receive it; silence leaves the question open.
This is not legal advice. Use this page to prepare questions for your counsel and the licensor.
Why outputs need separate answers
Robotics pipelines turn one licensed input into many outputs. Workplace video becomes 3D hand poses, then retargeted trajectories, then simulation assets, then synthetic video, then pretrained weights, then fine-tuned weights inside a product. A single "derivative works" sentence cannot answer all of those, and each step has a different owner, recipient and lifetime.
Derived data can also dwarf its seed. MimicGen generated more than 50,000 demonstrations from about 200 human ones (Mandlekar et al., 2023), and GR00T N1 reports that neural trajectories expanded teleoperation data from 88 hours to 827 hours (NVIDIA, 2025). If a license gives you "derivatives" but defines them narrowly, a synthetic expansion many times larger than the licensed seed may sit outside it.
The clause topics for the whole license are in licensing terms for robotics training data. This page goes deeper on what you may do with what you build.
Illustrative output-rights matrix
| Output | Example | Question to settle | What a buyer might ask for |
|---|---|---|---|
| Annotations and labels | Step segments, masks, language | Who owns labels made by the supplier, and who owns labels made by your own vendors? | Ownership or a perpetual license that survives the term |
| Derived measurements | 3D hand poses, object tracks, embeddings | Are they derivatives of the raw data? Can they be kept after the raw data is deleted? | A deletion scope that separates raw data from derived data |
| Transformed datasets | Cleaned clips, retargeted trajectories, converted formats | May they go to affiliates, collaborators or contractors? | Named recipient categories |
| Simulation assets | Meshes, textures, digital twins built from scans or CAD | May they be used in internal simulators? Inside a shared benchmark? | An express answer per asset class |
| Synthetic data | Generated video, neural trajectories, augmented episodes | Is a generated output a derivative? Who owns it? May it train other models? | Ownership, free use for training, survival after the term |
| Model weights | Pretrained, fine-tuned or distilled checkpoints | May weights ship in products, be hosted by a cloud provider, or be released openly? | Permission stated per distribution channel |
| Evaluation results | Benchmark scores, sample frames in papers | May you publish them? | A publication clause with review rights, if the licensor needs them |
Redistribution splits into four issues
- Raw data. Usually no redistribution. The practical issue is which processors may touch it: annotation vendors, cloud hosts and affiliates. Name recipient categories rather than "third parties", and see the glossary entry on sublicensing.
- Derived data. Collaborators, contractors and acquirers all raise different questions. Ask what happens to the license on a merger or change of control.
- Published artifacts. Sample frames, benchmark tasks and datasets released to the community can re-trigger obligations that the internal use did not.
- Weights. Hosting by a serving provider, embedding in robots sold to customers and open release are three different distributions.
Define each by recipient and purpose. A restriction written as "no redistribution" can be read to cover a contractor's laptop or to permit a customer's robot, depending on who is reading.
What a public license shows about the word "derivative"
Standardized licenses already show how much weight such words carry. Creative Commons states that NoDerivatives licenses bar public sharing of adaptations, that NonCommercial is judged by the use rather than the user, and that BY-NC-ND is its most restrictive license, allowing unadapted copies only (Creative Commons). It also states that CC licenses do not affect exceptions and limitations such as fair use.
The lesson for a bespoke robotics license is to define its own terms for annotations, trajectories, synthetic outputs and weights rather than borrow words whose meaning a court or counterparty may read differently. Whether a trained model counts as an adaptation under any given license is a legal question this page does not answer.
What the licensor can pass on
A licensor can grant only what it holds. If the supplier's material includes public datasets, vendor-made content or client goods, your derivative rights are bounded by those upstream terms. License labels on hosting platforms are also claims to verify: an audit of more than 1,800 text datasets found that, after re-annotation, 66% of the analyzed Hugging Face licenses fell into a different use category, often a more permissive one than the author's license (Longpre et al., 2023). That audit covers text, and its rate should not be assumed for robot data, but the method applies. For robot and egocentric datasets, see public dataset commercial license checks.
Copyright and AI training: what not to rely on
The U.S. Copyright Office released a pre-publication Part 3 report on generative AI training in May 2025. It identifies points in AI development where copying may implicate rights, analyzes fair use and examines licensing options (U.S. Copyright Office, 2025). It concerns generative AI, it is a pre-publication version, and it is not a ruling on training robot policies.
The practical consequence is to avoid depending on how those questions resolve. Ask for express permission for each stage of your pipeline and for each output, including collection, curation, pretraining, fine-tuning, evaluation and deployment.
Traps specific to robot pipelines
- Training-run ledger. Weights trained on a mix of licensed sources inherit questions from all of them. Keep a record of which dataset versions fed which checkpoint, or you cannot answer a scope, audit or deletion question later.
- Distillation and base models. A student model distilled from a licensed-data teacher, or a policy fine-tuned from a third-party base, may or may not be reached by the license. Raise it explicitly.
- Grant-back clauses. Check whether the licensor claims rights in your corrections, annotations or model outputs.
- Deletion and weights. Settle whether end-of-term deletion covers raw data only, or derived data and weights as well. See retention, deletion and audit rights.
- Non-exclusive grants. The licensor may license the same underlying data to others, including competitors; derived data you built should stay yours. Field limits and exclusivity are covered in exclusivity and field-of-use terms.
- Retargeted and synthetic outputs. Trajectories retargeted from human motion (retargeting human hand demonstrations) and generated data (real-world vs synthetic and simulation data) are the outputs most often left out of a license.
Illustrative output-rights schedule
# Illustrative only — synthetic parties; topics to raise with counsel, not model wording
licensor: supplier_S12
licensee: buyer_B03
raw_data: {redistribution: none, processors: named_categories_under_confidentiality}
outputs:
annotations_by_licensee: {owner: licensee, survives_term: true}
hand_poses_and_tracks: {owner: licensee, affected_by_raw_deletion: false}
retargeted_trajectories: {owner: licensee, share_with: [affiliates, named_collaborators]}
simulation_assets: {internal_simulators: permitted, redistribution: only_if_listed}
synthetic_data_from_licensed_data: {owner: licensee, may_train_other_models: true, survives_term: true}
model_weights: {ship_in_products: true, cloud_hosting: permitted, open_release: not_permitted}
evaluation_results: {publication: permitted, sample_frames: licensor_review}
assignment: notice_on_change_of_control
records_required: training_run_ledger # dataset versions per checkpoint
Questions to ask licensors
- Which outputs does the license name as permitted: annotations, poses, trajectories, simulation assets, synthetic data and weights?
- Who owns each output, and which of them survive termination?
- May model weights trained on the data ship inside products we sell, and be hosted by service providers?
- Which recipient categories may receive raw data and derived data, and what happens on a change of control?
- Does any upstream dataset, vendor content or client material limit what we can do with derivatives?
- May we publish benchmark results or sample frames, and with what review?
- Does end-of-term deletion reach derived data or weights?
- Does the licensor claim rights in our corrections, annotations or model outputs?
Discuss your requirements
SourceX agrees pricing and allowed uses in a license with the supplier, and nothing is contracted until a supplier agrees. Tell SourceX which outputs you need rights to together with the data requirement; it checks the data and the supplier's licensing permissions and follows up about potential sources. A request does not guarantee a match. SourceX's licensing terms glossary defines common terms, and the procurement hub and humanoid training data guide cover the rest of the sequence.
Explore sourcing and licensing options with SourceX
Sources
- About CC Licenses — meaning of NonCommercial and NoDerivatives, and the relation to fair use.
- Copyright and Artificial Intelligence, Part 3: Generative AI Training (pre-publication version) — US Copyright Office analysis of AI training, fair use and licensing options.
- The Data Provenance Initiative: A Large Scale Audit of Dataset Licensing & Attribution in AI — license labels that differ from authors' licenses.
- MimicGen: A Data Generation System for Scalable Robot Learning using Human Demonstrations — generated demonstrations from a small seed set.
- GR00T N1: An Open Foundation Model for Generalist Humanoid Robots — neural trajectories that expand teleoperation data.