Robotics and Autonomy IP Checklist: Component and Stack Mapping, Training Data Provenance, Safety Documentation, Integrator Terms, and Field Data Rights

By Casey Scott McKay · 2025-01-15

This checklist audits the intellectual property position of a robotics or autonomous systems business in the order the questions arise. It begins with the stack map, because no filing or protection decision can be made sensibly until each layer has a named owner and a named protection mode. It then works through the filing programme, the training data register, model weight controls, the safety documentation split, the fleet data clause, the integrator relationship, the simulation environment, the disclosure and export calendar, and the diligence file. Gate items mark the points at which work should stop until a specific artefact exists.

IP and Technology > Information Technology | Checklist | Published 15 January 2025 - Updated 28 July 2026 | Casey Scott McKay - marksy.us

Summary. This checklist audits the IP position of a robotics or autonomous systems business in the order the questions arise. It begins with the stack map, because no protection decision can be made sensibly until each layer has a named owner and protection mode. It then works through the filing programme, the training data register, model weight controls, the safety documentation split, the fleet data clause, the integrator relationship, the simulation environment, the disclosure and export calendar, and the diligence file. Gate items mark where work should stop until a specific artefact exists.

Keywords: robotics checklist · stack mapping · training data register · model weights control · safety case split · integrator work product · field data categories · teleoperation retention · simulation provenance · export classification · single actor claims · eligibility screening · demonstration discipline · aftermarket position · diligence file

How to use this checklist

| Phase | What it produces | Who runs it | Gate | |---|---|---|---| | 1. Stack map | One table naming every layer, owner, and protection mode | Counsel with engineering leads | Map exists and engineering has corrected it | | 2. Filing programme | A ranked list of what to file and how to draft it | Patent counsel | Nothing filed before the map | | 3. Training data | A register with one row per dataset | ML lead and counsel | No training run without a row | | 4. Model weights | Access controls plus a five-point contract wrapper | Security and counsel | No external copy without terms | | 5. Safety documentation | Two documents, distributed differently | Safety lead and counsel | Design rationale never sent unrestricted | | 6. Fleet data | A three-category clause | Commercial and counsel | Clause agreed before enterprise deals | | 7. Integrators | Work product, reporting, brand, and restriction terms | Partnerships and counsel | Integration work product resolved | | 8. Simulation | Provenance audit and access controls | Engineering and counsel | Bill of materials exists | | 9. Disclosure | A calendar and an export classification | Counsel | No public demonstration without review | | 10. Diligence | Ten artefacts an investor will ask for | Counsel | All ten exist as documents |

The matter. A warehouse robotics company has sixty machines deployed across six customers, three integrator partners, four patent applications all directed to the gripper, a perception stack the founders describe as the crown jewel, and a Series B process starting in five months. Nobody can name the datasets used to train the shipped model. The safety file has been sent in full to nine organisations. The integrator agreements assign integration work product to the integrators.

Phase 1. Build the stack map

[ ] List every layer of the system as a row: mechanical platform, sensors and sensing hardware, low-level control and calibration, safety architecture, perception, planning and decision, training and simulation apparatus, fleet management software, integration interfaces, and brand.
[ ] For each row, record who built it, who owns it, how it is protected today, and how it should be protected.
- Why. Every later decision references this table. A filing programme built without it aims at the layer the founders can describe rather than the layer that carries the value.
[ ] Verify the sensor layer position against the actual purchase terms, not against assumption. What the company owns in a bought sensor is the mounting, the calibration, and the fusion — not the sensor.
[ ] Find the calibration work and give it an owner. It sits between mechanical and software, it took years, it is why the machine is reliable, and it appears on no schedule in most companies.
- Trap. Work that spans two teams is work that neither team's manager reports, and it is systematically absent from invention harvesting.
[ ] Mark the layers currently protected by nothing. Usually perception, the training apparatus, and the simulation environment.
[ ] Circulate the map to engineering leadership and collect corrections. The corrections are the most valuable output of the exercise.
[ ] [Gate] No filing, licensing, or partnership decision proceeds until the map exists and has been corrected.

Layer-by-layer notes for the map

Mechanical platform. Usually the best-documented layer and the least defensible commercially, because there are many ways to build a chassis or a gripper. File on it anyway: it produces claims that read on articles a competitor sells and can be examined on a bench.

Sensors. Bought, almost always. Record the supplier, the purchase terms, the field-of-use restrictions, and whether the supplier has any right to the buyer's integration work. Note any supplier that is single-sourced and what qualifying an alternative would cost, because that number is the supplier's leverage in every future negotiation.

Low-level control and calibration. The layer most often unassigned to anyone. Ask specifically: who established the relationship between sensor frames and actuator frames, who tuned the loops, who wrote the procedure that a technician follows in the field. Those people are inventors and their work is patentable.

Safety architecture. Redundant sensing, envelope enforcement, safe stopping, fault detection and response. Two properties make this layer disproportionately valuable: it is technical enough to survive eligibility, and a competitor cannot design around it without changing what it certifies.

Perception. Where founders locate the value and where patent law helps least. Map what is model-based and what is engineered, because the engineered parts — preprocessing, geometric reasoning, sensor fusion structure — are often patentable even where the learned parts are not.

Planning and decision. Same analysis. The scheduling, arbitration, and fallback logic are frequently more claimable than the learned policy.

Training and simulation apparatus. Datasets, labelling pipeline, simulation environments, evaluation harnesses. Almost never on a schedule and almost always valuable.

Fleet management software. Where the recurring revenue lives, where the customer data flows, and where the terms of service govern more than the patent portfolio does.

Integration interfaces. See Phase 7. Record who wrote each one and what the governing agreement says about it.

Brand. Machine names, model designations, the company mark, and the domain portfolio. Cheap to protect and routinely neglected until a competitor files first.

Phase 2. Set the filing programme

[ ] File first on mechanical and electromechanical work: grippers, compliant mechanisms, tool changers, docking and charging, thermal management, cable routing through articulated joints. Classical, patentable, and enforceable against an article a competitor sells.
[ ] File second on calibration, sensing methods, and the safety architecture. Both recite improvements to machine operation, which is the framing most likely to survive the Alice Corp. v. CLS Bank International analysis under 35 U.S.C. § 101.
[ ] File third on integration and interface methods, which read on the deployment rather than on the competing robot and therefore identify both an infringing act and a servable defendant.
[ ] File selectively inside perception on specific mechanisms, never on capabilities. A particular data structure or a named failure-mode handling, not "identifying obstacles more accurately."
[ ] Draft apparatus claims reciting the machine wherever possible, following the reasoning in Diamond v. Diehr that a process controlling a physical operation is not unpatentable merely because it uses an algorithm.
[ ] Confine every claim to a single actor. Where the system distributes work across machine, cloud, and human operator, write separate claim sets, because Limelight Networks, Inc. v. Akamai Technologies, Inc. defeats enforcement of a claim no one entity performs.
[ ] Do not claim anything whose operation depends on an undisclosed training corpus. 35 U.S.C. § 112 and Amgen Inc. v. Sanofi mean such a claim is not enabled across its scope.
[ ] Cover both the autonomous and the supervised configuration where the product ships with teleoperation available.
[ ] Record human contribution to conception where design tooling generated candidates, since Thaler v. Vidal confirms an inventor must be a natural person and 35 U.S.C. § 115 requires the declaration.
[ ] [Gate] No application is filed on a perception capability rather than a perception mechanism.

Phase 3. Build the training data register

[ ] Create one row per dataset used in any training run contributing to a shipped model, with source, acquisition date, licence, permitted uses, restrictions, commercial-use permission, derivative-model permission, approver, and the model versions affected.
[ ] Populate it from three directions: ask the machine learning team, search the training infrastructure for data paths, and read the shipping model's training configuration. The three answers will differ.
[ ] Flag research and academic benchmarks with non-commercial or research-only terms, which is the most common fatal finding.
[ ] Flag asset store content used in simulation, usually licensed per-project and frequently not licensed for machine learning at all.
[ ] Assess scraped content on both the copyright question — where Feist Publications, Inc. v. Rural Telephone Service Co. establishes only that the underlying facts are free — and the contractual question of whether collection breached terms of use.
[ ] Check customer data against the scope the customer agreement permitted, including whether model training was contemplated at all.
[ ] Check employee-contributed data collected before assignment terms were in place.
[ ] For each finding, choose one of three options and record it: retrain without it, obtain a licence, or accept and document the risk.
- Trap. Accepting risk is legitimate. Accepting it without knowing is what produces a valuation discount, because a buyer assumes the worst case for anything undocumented.
[ ] [Gate] No dataset enters a training run without a register row and a named approver.

Reading a training data licence

Six questions decide whether a dataset can be used, and a register row that does not answer all six is incomplete.

Is commercial use permitted at all? Research-only and non-commercial terms are the most common blocker, and they appear on datasets that everyone in the field uses, which creates a false sense of safety. Widespread breach is not a defence.

Is machine learning use contemplated? Older licences predate the practice and say nothing. Silence is not permission where the licence grants specified rights and reserves the rest, and it is closer to permission where the licence grants broad use subject to named restrictions. Read which structure you have.

Are derivative models permitted, and are they encumbered? Some terms purport to attach conditions to anything trained on the data — attribution, share-alike, or field restrictions. A model carrying such an encumbrance is a model whose commercial deployment may breach the licence.

Does the licence survive? Terms that terminate on the licensor's insolvency, on acquisition, or at will leave a shipped model resting on a licence that may not exist next year.

Who granted it, and did they have the right? Aggregated datasets frequently combine material the aggregator did not own. A licence from someone without rights transfers nothing, and the register should record whether the chain was checked.

What are the attribution and notice duties? Cheap to comply with and embarrassing to breach, and several widely used corpora impose them.

[ ] Answer all six for every row.
[ ] Re-check on any licence version change, since dataset terms are revised more often than software licences.
[ ] Record the answer, not the conclusion. "Permitted" is not a record; "commercial use permitted, ML use not addressed, derivative models unrestricted, perpetual, chain verified, attribution required" is.

Phase 4. Control the model weights

[ ] Stop trying to resolve whether trained weights are a copyrightable work. The question is open and the protections that work do not depend on it.
[ ] Restrict access to named individuals with a business need, and maintain the list.
[ ] Encrypt at rest, mark every artefact, and log access. These are the measures that make the reasonable measures element of 18 U.S.C. § 1839 provable rather than assertable.
[ ] Confirm offboarding actually revokes access to model storage, training infrastructure, and any copies on personal devices.
[ ] Apply the five-point contract wrapper to every external copy: permitted use stated affirmatively; express prohibition on extraction, distillation, and architecture reverse engineering; treatment of outputs including whether they may train anything; concrete security obligations; and a defined fate on termination with certification of deletion.
- Why. Kewanee Oil Co. v. Bicron Corp. confirms that choosing secrecy over patenting is legitimate, but secrecy is only as good as the terms that travel with each disclosure.
[ ] Where the model ships to customer hardware, add technical measures: encryption, secure enclaves, integrity checks. Note that access controls engage 17 U.S.C. § 1201, which gives a cause of action against circumvention while drawing the product into a regime with a periodic exemption process.
[ ] Where the model is served from your infrastructure, defend against query extraction with rate limits, query logging, anomaly detection, and terms of service prohibiting the conduct.
[ ] [Gate] No model leaves the company's control without the five-point wrapper in the agreement covering it.

Phase 5. Split the safety documentation

[ ] Separate the safety case from the design rationale. The safety case describes behaviour: what the system does under specified conditions, its limits, fault detection and response, residual risk, and operator duties. The design rationale describes implementation: sensor fusion approach, calibration procedure, model architecture, rejected alternatives, component-level failure analysis.
[ ] Distribute the safety case freely and the design rationale narrowly, under terms specific to it, referenced rather than reproduced in the safety case.
- Why. The safety file is the most complete description of the system that will ever be written, and it goes to customers, integrators, insurers, regulators, and eventually opposing counsel. A competitor who obtained it would learn more than from any amount of reverse engineering.
[ ] Claim confidentiality at submission where material goes to a regulator, with specificity rather than as a blanket assertion, because a claim not made at submission is often not available later.
[ ] Number copies, name recipients, and maintain a register. A document sent to nine organisations without terms is nine uncontrolled copies.
[ ] Bind the recipient's contractors in the terms, since a customer's safety review routinely involves external consultants.
[ ] Write assuming litigation discovery. After an incident the safety file is requested first, protective orders manage rather than eliminate exposure, and appellate courts have grown less tolerant of blanket sealing designations.
[ ] [Gate] The design rationale is never sent without document-specific terms and a register entry.

Phase 6. Rewrite the fleet data clause into three categories

[ ] Concede operational data — what was moved, how many, when, for whom. It describes the customer's business, and fighting for it costs credibility needed elsewhere.
[ ] Take broad rights to machine data — sensor readings, control outputs, fault conditions, intervention events, calibration drift, component wear — as an irrevocable, perpetual, worldwide licence for any purpose relating to the vendor's products and services.
[ ] Own derived models and improvements outright, with the customer's benefit delivered through product improvement rather than ownership of a model it cannot use.
[ ] Address incidental operational content in machine data: camera logs record the facility. Commit to named, commercially reasonable measures to avoid retaining identifiable operational content.
[ ] Resolve the competitor restriction with a term limiting use to general product development rather than bespoke work for a named competitor.
[ ] State survival expressly for data already collected.
- Trap. A vendor obliged to purge its corpus when a customer departs has built its model on borrowed material, and the obligation is easy to agree to and impossible to unwind.
[ ] Handle personal data in a separate provision, since camera and audio data captured around workers may be personal data and in several states biometric data with statutory consent requirements.
[ ] Answer the security and access questions specifically: where the data goes, who can see it, and for how long.
[ ] [Gate] The three-category clause is settled before the first enterprise deal, not negotiated inside it.

What the deployed fleet is telling you

[ ] Confirm you actually receive the machine data your agreements entitle you to. Vendors frequently negotiate broad data rights and then never build the pipeline, which means the clause is a right without an asset behind it.
[ ] Check whether integrators are intercepting it. Where the integrator operates the fleet management layer, machine data may terminate at the integrator and never reach the vendor at all.
[ ] Verify the anonymisation measures you promised are implemented, since a commitment to commercially reasonable measures that engineering never received is a breach waiting for an audit.
[ ] Reconcile what you collect against what the customer agreement permits, in both directions: data collected beyond the grant is a problem, and data the grant permits but the pipeline discards is a wasted asset.

Phase 7. Fix the integrator relationship

[ ] Resolve integration work product first. The interface between the robot and the customer's systems is built by the integrator and, absent a term, owned by the integrator — and that interface is what makes the machine deployable.
- Why. An integrator who owns the interface can put a competing robot behind it, and the customer will not notice.
[ ] Obtain assignment, or at minimum a broad, perpetual, sublicensable licence covering reuse of the interface with other customers.
[ ] Impose a field reporting obligation: commissioning reports, fault reports, and field data returned to the vendor as a condition of the partnership.
- Trap. The vendor loses visibility of its own product's behaviour, which is exactly the data the next model needs.
[ ] Secure brand and contact terms: vendor branding visible on the machine and in the interface, vendor named in warranty documentation, and a right of direct contact for safety and support matters.
[ ] Restrict use of vendor-supplied training, documentation, and tooling to support competing products.
[ ] Mark, limit, and bind the technical disclosures — interface specifications, tuning guidance, failure documentation, and source where provided — including the integrator's own subcontractors, with return or destruction at termination.
[ ] Decide exclusivity deliberately. It usually costs more than it is worth, and occasionally it is worth it.
[ ] [Gate] No new integrator agreement is signed without the integration work product term.

Phase 8. Audit the simulation environment

[ ] Protect it as an asset. Access control, marking, and the same trade secret discipline applied to the models. It determines development speed and embeds years of knowledge about what physically matters.
[ ] Produce a bill of materials covering physics and rendering engines, asset store content, customer facility models, adapted research scenarios, and third-party CAD models.
[ ] Check physics and rendering engine licences for commercial-use restrictions, attribution requirements, and copyleft obligations on distributed builds.
[ ] Check asset store terms, which are frequently per-seat or per-project and often silent or prohibitive on machine learning use.
[ ] Identify customer facility models supplied under confidentiality for a specific deployment and later retained in the general library.
- Trap. This is a confidentiality breach hiding inside a technical convenience, and it is the finding most likely to damage a customer relationship.
[ ] Determine whether any part of the environment is distributed — for acceptance testing, commissioning, or research partnership — because distribution triggers whatever obligations the components carry.
[ ] [Gate] No simulation content is distributed externally without a current bill of materials.

Standards exposure, in one pass

[ ] List every standards body the company participates in and every contribution it has made.
- Why. Standards organisations typically require participants to disclose essential patents and to license them on defined terms, and those commitments run with the patents and bind assignees. A company that contributed to an interoperability standard three years ago may hold encumbered assets nobody recorded.
[ ] Check whether connectivity modules carry a licence covering the end product or only the chip, since that determines whether the company has standard-essential patent exposure it has not priced.
[ ] Run freedom to operate on the safety-critical path first. A third-party claim reading on the only practical means of satisfying a mandatory safety requirement cannot be designed around without failing certification, which makes it the most expensive kind of blocking right and the one least often searched for.
[ ] Record any commitment the company has made in the encumbrance list that goes to Phase 12.

Aftermarket position, decided deliberately

[ ] Establish what the law permits. A purchaser may repair; reconstruction is different, and Aro Manufacturing Co. v. Convertible Top Replacement Co. remains the reference. An authorised sale exhausts the patent right regardless of post-sale conditions per Impression Products, Inc. v. Lexmark International, Inc., so restrictions on parts and service run in contract or not at all.
[ ] Understand what access controls give you. Gating diagnostics engages 17 U.S.C. § 1201, which supplies a cause of action but sits inside a triennial exemption process that has repeatedly considered device repair.
[ ] Build the position architecturally where possible. Calibration dependency after a component change is more durable than a contractual restriction, and it is decided by service architecture choices made early — which is why counsel should be present when that architecture is designed.
[ ] Price the service business honestly, because aggressive lockout produces customer resentment, regulatory attention, and legislative interest that a fair-terms position does not.

Phase 9. Run the disclosure and export calendar

[ ] Catalogue the disclosure events for the coming year: trade shows, customer pilots, investor demonstrations, published videos, conference papers, and standards contributions.
[ ] Route each through three questions: does it disclose a claimable invention, does it disclose a trade secret, and is it an export.
[ ] File before any public demonstration of a new capability. Under 35 U.S.C. § 102 a public demonstration of the claimed method can start the domestic grace period and can be an outright bar where no grace period exists.
- Why. The grace period is a rescue, not a plan, and most commercially important jurisdictions have none.
[ ] Paper every pilot. A pilot under confidentiality is not a public use, but the agreement must exist, must cover the subject matter, and must bind the customer's contractors and visitors.
[ ] Write down the export classification for the autonomy stack, the sensing hardware, and the technical data, item by item.
[ ] Map deemed export exposure: distributed engineering teams, foreign investors conducting technical diligence, international research collaborators, and foreign patent associates.
[ ] Check every foreign filing for a licence under 35 U.S.C. § 184, since 35 U.S.C. § 185 makes an unlicensed foreign filing a ground of invalidity.
[ ] Screen for secrecy order exposure under 35 U.S.C. § 181 where the technology has defence applications.
[ ] [Gate] No public demonstration, publication, or foreign filing without a completed review record.

Phase 10. Set the teleoperation retention policy

[ ] Recognise what the record is. Every teleoperation session captures when the machine failed, what a human did instead, and what the sensors showed. It is the most valuable training data the company holds and the most damaging discovery material it will produce.
[ ] Set the policy before any incident, covering retention period, form, access controls, and deletion authority.
- Trap. A retention policy adopted before an incident is a business record. A policy changed after one is a fact a plaintiff will develop at length.
[ ] Write down the reasoning, because the reasoning is what makes the policy defensible.
[ ] File on the teleoperation interface — situational awareness presentation, latency handling, control handover, degraded-link behaviour — which is patentable and routinely overlooked as a product feature.
[ ] Handle operator monitoring obligations, since teleoperators are workers recorded continuously and several jurisdictions impose notice and consent duties.
[ ] Secure ownership of operator-generated annotations and corrections, which become training labels, especially where operators are contracted through a third party.
[ ] [Gate] The retention policy is written and approved before the next deployment expansion.

Phase 11. Employment and departure controls

[ ] Confirm every employee and contractor agreement contains a present assignment, not a promise to assign.
- Trap. Contractor paper is where assignment language is most often missing, and robotics companies use specialist contractors for mechanical design, safety assessment, and machine learning work.
[ ] Name the confidentiality categories specifically: model weights, training data, calibration procedures, simulation assets, design rationale, and field failure data.
[ ] Run a structured exit: access revocation on the day, device return, a documented interview walking through the categories, and a forensic snapshot where circumstances warrant.
[ ] Screen incoming hires for prior obligations and obtain a written acknowledgement that they will not bring or use a former employer's confidential material.
[ ] Keep contemporaneous inventorship notes where design tooling contributed to a claimed invention.
[ ] [Gate] No departure from an engineering role proceeds without the exit process.

Phase 12. Assemble the diligence file

[ ] Produce the ten artefacts an investor or acquirer will ask for: the stack map; the training data register; the export classification; the model access and control record; the safety documentation structure with its distribution register; the current fleet data clause; the integrator agreements with the work product position highlighted; the simulation bill of materials; the disclosure review log; and the chain of title for every application.
[ ] Support each answer with a document, not a memorandum summarising documents that do not exist.
[ ] Stage data room access so controlled technical data is not disclosed to an unscreened reviewer, since foreign investment screening examines technology transfer specifically.
[ ] List every encumbrance: standards commitments, government funding obligations under 35 U.S.C. § 202, field restrictions, exclusivity commitments, and security interests.
[ ] Reconcile the freedom-to-operate position on the safety-critical path, which is where a blocking third-party claim would be hardest to design around.
[ ] Confirm no application is subject to an undisclosed secrecy order.
[ ] [Gate] The file is complete when all ten artefacts exist as documents.

A note on order

The phases are ordered by dependency, not by importance. The stack map comes first because every later decision references it — a filing programme without it aims at whichever layer the founders described most fluently, and a diligence file without it is a list of assertions.

The training register comes early because remediation takes engineering time rather than legal time. A dataset that has to be removed means a retraining run, an evaluation cycle, and a release, and none of those compresses. Discovering the problem in month one of a five-month process is manageable; discovering it in month four is not.

The safety split comes before the integrator work because the document that has already been distributed cannot be recalled, and every week of delay adds recipients. The fleet data clause comes before the first enterprise negotiation for the same reason in reverse: a position conceded once becomes the precedent every later customer cites.

The integrator conversations come late in the sequence and take longest, because they are commercial rather than legal. A vendor asking three partners to renegotiate the term that gives those partners their leverage is asking for something, and the ask needs a trade. Starting the conversation in month three of a five-month process means it is underway rather than concluded when diligence begins, which is an acceptable answer to an investor who wants to see that the problem is identified.

The diligence file comes last because it is a compilation. Every artefact in it is the output of an earlier phase.

Outcome. A business that has run this checklist can say which layer of its system carries the value, whether the model it ships was trained on material it was licensed to use, who owns the interface that makes the product deployable, and whether its most detailed technical description is sitting unrestricted in nine other organisations. Those four answers determine what the company is worth, and none of them appears in the technical file.

Key Authorities at a Glance

| Authority | What it settles | Phase | |---|---|---| | 35 U.S.C. § 101 | Patentable subject matter | 2 | | Alice Corp. v. CLS Bank International | Two-step abstract idea framework | 2 | | Diamond v. Diehr | A process controlling a physical operation is eligible | 2 | | Bilski v. Kappos | Machine-or-transformation is a clue, not the test | 2 | | 35 U.S.C. § 112 | Written description and enablement | 2 | | Amgen Inc. v. Sanofi | Functional breadth must be enabled | 2 | | Limelight Networks, Inc. v. Akamai Technologies, Inc. | Single-actor requirement for direct infringement | 2 | | 35 U.S.C. § 271 | Acts of infringement | 2, 7 | | Thaler v. Vidal | An inventor must be a natural person | 2, 11 | | 35 U.S.C. § 115 | Inventor's oath or declaration | 2, 11 | | 18 U.S.C. § 1839 | Reasonable measures element of trade secret status | 4, 7, 8, 11 | | Kewanee Oil Co. v. Bicron Corp. | Trade secret law coexists with patent law | 4 | | Feist Publications, Inc. v. Rural Telephone Service Co. | Facts are unprotectable | 3 | | 17 U.S.C. § 1201 | Circumvention and triennial exemptions | 4 | | 35 U.S.C. § 102 | Novelty, public use, grace period | 9 | | 35 U.S.C. § 184 and § 185 | Foreign filing licence and invalidity | 9 | | 35 U.S.C. § 181 | Secrecy orders | 9 | | 35 U.S.C. § 202 | Bayh-Dole obligations | 12 | | Aro Manufacturing Co. v. Convertible Top Replacement Co. | Repair permitted, reconstruction not | Aftermarket position | | Impression Products, Inc. v. Lexmark International, Inc. | Authorised sale exhausts the patent right | Aftermarket position |

The five things people get wrong

One: filing on the capability instead of the mechanism. Founders describe what the machine does, and the instinct is to claim that. A claim to identifying objects and selecting actions is a claim to gathering information, analysing it, and acting — which is where the Alice analysis ends badly. The claims that survive recite specific technical mechanisms, and the claims that are worth having recite mechanisms a competitor cannot avoid: calibration, safety architecture, and the interface that makes the product deployable.

Two: not knowing what the model was trained on. Almost no robotics company can produce a complete dataset list on request, and almost every register that gets built surfaces at least one dataset whose terms prohibit what the company is doing. The finding is cheap to remediate at the time of discovery — a retraining run — and expensive to remediate during diligence, where it becomes a price adjustment. Build the register before someone else asks for it.

Three: sending the whole safety file. The safety documentation is written by safety engineers for safety engineers, and it is complete by design. Sent in full to every customer and integrator, it distributes the most detailed description of the system that exists, without terms, to organisations that will keep it after the relationship ends. Split it: a behavioural safety case that goes everywhere, and a design rationale that goes nowhere without specific terms and a register entry.

Four: letting the integrator own the integration. The interface between the robot and the customer's systems is what makes the machine sellable, and it is built by a partner who owns it absent a term to the contrary. Three years later the partner has a generic interface, a qualified alternative machine, and no reason to keep buying yours. This is the single most consequential term in a robotics company's contract stack and it is routinely left to the partner's template.

Five: treating the fleet data clause as boilerplate. Data from deployed machines is what improves the next model, and the clause that governs it is usually copied from a software agreement that never contemplated a camera in a warehouse. Three categories fix it — operational data conceded, machine data taken broadly with survival, derived models owned outright — and the fix has to happen before the first enterprise negotiation, because a customer who has been given ownership once will expect it again.