Due to our team’s recent breakthroughs in physical AI and AIoT, this website and the websites of all 64 portfolio companies are being substantially upgraded to incorporate our Physical AI and AIoT engines and sensing technologies. This overhaul will be completed by Oct. 31, 2026.
This page has not been updated yet.

Industrial Knowledge Learning and Deployment

This research program is conducted by a team of Ph.D. researchers with expertise in knowledge learning and deployment.
Turn research prototypes into maintainable industrial capabilities through reliable context, efficient learning and reproducible evaluation. This program addresses practical reuse without assuming identical assets or operating conditions.

Industrial Knowledge Integration Asset Administration Shells and Semantic Interoperability

Align asset identities, units, timestamps, work orders and process terminology across ERP, MES, WMS, CMMS, historians and device interfaces. Evaluate typed data contracts, knowledge graphs and retrieval supported explanations. Preserve provenance and conflicting evidence rather than silently merging inconsistent records. Evaluate Asset Administration Shells (AAS) and selected submodels as candidate structures for asset descriptions, capabilities and references to telemetry and AI model metadata. Extend digital threads across product lifecycle management (PLM), MES and ERP. Validate units, semantic identifiers, versions and access rights. Common structures can support exchange but do not remove the need to resolve missing or incompatible meaning.

Research questions

How can an assistant retrieve the correct procedure for an exact equipment version? Which semantic checks prevent a valid value from being applied to the wrong asset or unit? How much mapping can be automated reliably, and which ambiguous fields require expert review?

Proposed Demonstration

Integrate two deliberately inconsistent equipment datasets and operating manuals, then answer traceable maintenance and status queries. Compare a manually mapped baseline with AAS based exchange across two equipment sources, including conflicting units, incomplete metadata and changed schema versions.

Evaluation

Entity matching accuracy, unit and version errors, evidence coverage, unsupported answers and integration effort. Measure mapping hours, expert corrections, schema conformance, digital thread completeness and portability across vendors.

Learning from Limited Labels and Industrial Expert Feedback

Compare transfer learning, active learning, self supervised pretraining and expert annotation for rare faults and defects. Study uncertain and inconsistent labels, site variation and the cost of expert review. Evaluate foundation models as candidate components against simpler baselines.

Research questions

Which examples should an expert label first? How much benefit remains when the evaluation uses an unseen site, machine or defect category?

Proposed demonstration

Build learning curves for a defined inspection or diagnostic task with staged labeling budgets and independently reviewed test labels.

Evaluation

Performance at each label budget, expert hours, calibration, cross site degradation and sensitivity to label disagreement.

Industrial Model Monitoring Adaptation and Controlled Release

Monitor changes in input quality, operating conditions, prediction confidence and realized outcomes after deployment. Separate process drift from failures of the model or data pipeline. Evaluate shadow operation, staged rollout and rollback; retraining should produce a reviewed version rather than an uncontrolled live change.

Research questions

How can performance decline be detected when ground truth arrives late? What evidence is sufficient to promote a new model without hiding rare failure regressions?

Proposed demonstration

Replay time ordered equipment or inspection data through old and candidate models, with simulated drift and delayed labels.

Evaluation

Drift detection delay, false triggers, regression detection, retained task performance and rollback effectiveness.

Reusable Engine Interfaces and End to End Industrial Benchmarks

Define shared records for identity, measurements, uncertainty, decisions, approvals, actions and verified outcomes. Evaluate what can transfer between the Aperture AIoT Engine™ and Aperture Physical AI Engine™ implementations while preserving application specific constraints. Reuse should be measured through integration effort and maintained performance. Include platform adapters for robots, drones and autonomous vehicles, and carry data integrity, permission scope and command freshness through the shared records.

Research questions

Which interface contracts remain stable across inventory and equipment health applications? Does reuse reduce deployment effort without concealing incompatible units, timing or action semantics?

Proposed demonstration

Implement one tracking workflow and one condition monitoring workflow with shared components; add bounded physical action to one and test feedback end to end.

Evaluation

Integration time, reusable component coverage, interface errors, end to end task success, feedback completeness and adaptation effort.

Operator Intervention Capture and Learning from Teleoperation

Treat operator corrections, remote takeovers, rejected recommendations and failed actions as structured learning evidence. Investigate capture of the surrounding state, the reason for intervention and the corrective action taken. Study labelling of inconsistent or partially justified interventions and selection of examples that improve policies without reinforcing operator error. Route resulting model changes through the reviewed release process of Topic 57 and protect intervention channels as described in Topic 51.

Research questions

Which interventions contain information that improves future behavior, and which reflect operator preference, caution or error? How many intervention examples are needed to reduce the rate of similar interventions without increasing constraint violations?

Proposed demonstration

Operate a manipulation or mobile delivery task under supervision, recording interventions over repeated sessions. Retrain a policy using curated intervention data, release it in shadow operation and compare intervention rates with the original policy.

Evaluation

Intervention rate over time, recurrence of previously corrected failures, constraint violations after retraining, labelling effort, proportion of interventions retained as useful training data and operator workload.

Federated and Cross Site Learning Across Customer Deployments

Investigate how models for diagnostics, inspection and identification can improve across multiple customer sites without pooling raw operational data. Compare federated learning, shared pretrained components with local adaptation and exchange of aggregate statistics. Address heterogeneous equipment, operating conditions, label definitions and data volume between sites. Measure the risk that shared model updates disclose proprietary process information or allow reconstruction of individual records. Require customer permission and defined data rights. Treat federated learning as a data sharing architecture, not a privacy guarantee; evaluate secure aggregation and differential privacy where suitable, including their utility costs and exposure to poisoned updates.

Research questions

When does cross site learning outperform independent models trained at each site? Can a site that differs substantially from others benefit from shared learning without degrading performance elsewhere?

Proposed demonstration

Partition a predictive maintenance or inspection dataset into simulated customer sites with differing equipment and conditions. Compare local training, centralized training and federated approaches, then conduct membership inference and reconstruction attempts against shared updates.

Evaluation

Performance at each site, benefit for small and atypical sites, communication and compute cost, convergence stability, measured information leakage and effort to onboard a new site.

Techno Economic Evaluation and Deployment Value Evidence

Develop repeatable methods for estimating and measuring the economic value of AIoT and physical AI deployments. Include hardware, integration, data preparation, validation, training, maintenance and supervision costs alongside operational benefits such as reduced downtime, inventory error, scrap, energy use and manual effort. Compare predicted benefits with realized results and identify the assumptions responsible for divergence. Keep economic conclusions traceable to measured operating data and explicit assumptions.

Research questions

Which cost components are most frequently underestimated in industrial AI deployments? How accurately can pilot results predict value at full deployment scale, and which operating conditions cause that prediction to fail?

Proposed demonstration

Define costs, expected benefits and baseline methods before two selected pilots enter extended operation. Compare predictions with subsequently measured outcomes, normalize for workload and operating conditions, and report uncertainty. Use retrospective pilots only as a separately labelled exploratory comparison.

Evaluation

Accuracy of predicted versus realized benefit, completeness of captured costs, integration hours, time to measurable value, sensitivity to key assumptions and consistency of the method across applications.

Explore the Other Programs