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 Nov. 30, 2026.
This page has not been updated yet.

Industrial Digital Twins and Simulation

Use calibrated models and repeatable experiments to evaluate decisions before broader physical deployment. Focus on prediction quality and decision usefulness rather than visual representation alone.

This research program is conducted by a team of Ph.D. researchers with expertise in digital twins.

Operational Digital Twins for State Estimation and Action Evaluation

Combine physics based models and live measurements to estimate hidden process states and compare candidate actions. Study synchronization, parameter identification and drift as equipment ages. Represent uncertainty in twin predictions and define when the twin is unsuitable for a decision.

Research questions

How accurate must a model be for a particular action? Can online calibration distinguish changed equipment behavior from sensor faults?

Proposed Demonstration

Build a pump, motor or thermal twin and compare predicted action responses with independent physical measurements.

Evaluation

Prediction error, synchronization delay, uncertainty calibration, drift detection and improvement in decisions compared with a sensor only baseline.

Simulation to Reality Transfer and Synthetic Industrial Data

Use simulated variation to train or stress test inspection, perception and control methods. Investigate domain randomization, realistic defect generation and calibration with small physical datasets. Keep synthetic and physical test sets separate and assess whether simulated difficulty matches field conditions. Extend selected scenarios to robot manipulation, drone flight and autonomous vehicle navigation, including rare obstacles, changing terrain and unreliable localization.

Research questions

Which variations improve transfer and which introduce misleading shortcuts? Can the system detect operation outside the simulator’s assumptions?

Proposed Demonstration

Train an inspection or bounded control model with mixed synthetic and real data; test on untouched physical examples and unusual conditions.

Evaluation

Real world performance, physical data needed, transfer degradation, rare case coverage and sensitivity to simulator mismatch.

Hardware in the Loop Testing and Scenario Coverage

Connect real controllers or gateways to simulated processes so timing, protocol behavior and failure handling can be tested repeatably. Build scenarios for stale data, saturation, communication loss and conflicting commands. Measure test coverage against defined hazards and operational requirements.

Research questions

Which faults require real hardware to expose timing failures? How can a test suite reveal combinations of individually tolerable faults that cause unsafe behavior?

Proposed Demonstration

Run a gateway and controller against a process simulator, replaying a versioned set of nominal and fault scenarios.

Evaluation

Scenario coverage, detected integration faults, timing violations, repeatability and agreement between simulated and physical rig results.

Counterfactual Process Experiments and Production Flow Twins

Model queues, resource contention, changeovers and material availability to compare production policies. Combine discrete event simulation with measured processing times and uncertainty. Use counterfactual comparisons as decision hypotheses that require operational validation.

Research questions

Which bottlenecks persist after demand or staffing changes? Can a scheduling improvement remain useful when cycle times and equipment availability are uncertain?

Proposed Demonstration

Calibrate a small production flow twin from event logs; compare proposed schedules with historical replay and a limited supervised trial.

Evaluation

Queue and cycle time prediction, throughput, work in progress, schedule feasibility and realized improvement against the predicted benefit.

Industrial World Models and Action Conditioned Prediction

Investigate learned models that predict the physical consequences of candidate actions from sensor histories, video and process measurements. Compare learned world models with physics based twins and study hybrid approaches in which physical models constrain learned prediction. Address prediction horizon, compounding error, rare events and the representation of uncertainty. Define criteria for recognizing when a learned prediction falls outside its validated range and must not be used to select an action.

Research questions

For which industrial tasks does a learned world model predict action outcomes more usefully than a calibrated physics model? How quickly does prediction quality degrade with horizon length, and can the model reliably signal that degradation?

Proposed Demonstration

Train an action conditioned model on a conveyor sorting, thermal or manipulation task using physical and simulated data. Evaluate predicted outcomes of candidate actions against measured results, including unfamiliar objects and disturbances, and compare with the operational twin of Topic 20.

Evaluation

Prediction error by horizon, uncertainty calibration, out of distribution detection, decision quality when predictions select actions, data requirements and computational cost relative to physics based and hybrid models.

Explore the Other Programs