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.

Verified AI Decisions and Industrial Control

Translate operational intent into constrained decisions and measurable physical responses. Selected work connects the AI Decision Engine with the Physical AI Action Engine through checked command interfaces and outcome feedback.
This research program is conducted by a team of Ph.D. researchers with expertise in AI decisions and physical control.

Verified AI Agents for Industrial Device and Equipment Control

Translate operational requests into a bounded action vocabulary linked to equipment identities and current state. Investigate structured planning, precondition checks, authorization, execution monitoring and evidence of completion. Separate interpreting a request from granting permission to execute it.

Research questions

Can an agent resolve an ambiguous target safely? How can it recognize that a command was acknowledged but the intended physical change did not occur?

Proposed Demonstration

Control a pump, fan or conveyor through a restricted gateway; introduce ambiguous requests, actuator failures and delayed feedback.

Evaluation

Task completion, invalid proposals, unauthorized commands blocked, constraint violations, physical verification accuracy and recovery time.

Constraint Aware Adaptive Control and Industrial Optimization

Evaluate supervisory setpoints and schedules using learned process models, physical constraints and uncertainty estimates. Compare conventional control with model predictive and learning assisted approaches. Include limits on temperature, pressure, rate of change, equipment wear and product quality.

Research questions

How should uncertainty shrink the available action range? When does adaptation improve performance without destabilizing the underlying control loop?

Proposed demonstration

Optimize a pump or thermal rig under variable demand and disturbances while an independent controller enforces defined limits.

Evaluation

Tracking error, resource use, throughput, constraint violations, recovery after disturbance and performance outside the training regime.

AIoT and Physical AI for Building and Energy System Operation

Optimize a defined subsystem such as ventilation, cooling, compressed air or thermal storage using occupancy, weather, tariffs and equipment condition. Investigate interactions between demand prediction, comfort or process requirements and permitted operating schedules.

Research questions

Can demand response reduce peaks without shifting excessive consumption elsewhere? How should missing occupancy data or forecast error alter operating decisions?

Proposed demonstration

Use a ventilation testbed or supervised facilities pilot with a weather and occupancy normalized baseline.

Evaluation

Energy use, peak demand, comfort or process compliance, equipment cycling, operator interventions and robustness to forecast error.

Cold Chain Intelligence and Verified Environmental Intervention

Associate product identity and custody with temperature, humidity, door openings and transport conditions. Study cumulative excursion exposure, thermal lag and risk forecasting. Evaluate approved refrigeration adjustments, rerouting recommendations or quarantine workflows; product disposition must use validated product specific criteria.

Research questions

Can the model forecast an excursion early enough to intervene? How can it distinguish brief air temperature changes from sustained product temperature risk?

Proposed demonstration

Instrument a cold box with reference product probes; vary loading, door openings and cooling interruptions, then test bounded corrective actions.

Evaluation

Excursion warning lead time, product temperature prediction, false quarantine recommendations, recovery time and energy use.

Water Process Leak and Environmental Control Intelligence

Combine pressure, flow, level and water quality measurements to identify leaks, abnormal consumption and process deviations. Study network state estimation, pump scheduling and constrained dosing decisions where process expertise is available. Account for transport delays and sensor fouling.

Research questions

Can leakage be distinguished from legitimate demand variation? How should uncertain quality measurements restrict automatic intervention?

Proposed demonstration

Use a recirculating water rig or validated simulation with controlled leaks, changing demand and delayed measurements before any supervised field trial.

Evaluation

Leak detection delay, localization error, false alarms, water and energy use, process compliance and safe response under uncertain data.

Agentic AIoT for Multistep Planning and Long Horizon Operations

Investigate stateful industrial agents that maintain task memory, use approved tools and revise multistep plans as physical conditions change. Separate planning from command authority and fast local control. Revalidate asset identity, observation freshness, permissions and physical preconditions at each action; use bounded retries, checkpoints and explicit stop or escalation rules. Self reflection is not verification, and reversing a software transaction does not reverse every physical action.

Research questions

Can an agent recover a partially completed workflow without repeating an irreversible action? How should changing state invalidate earlier approvals and require replanning or operator review?

Proposed demonstration

Use a simulated multistation line or bounded laboratory rig for staged recovery after an outage. Introduce moved assets, partial completion, stale memory and unavailable equipment. Compare a fixed workflow, an agentic planner and an agent with independent action checks.

Evaluation

Verified end to end completion, sequencing errors, stale approvals rejected, duplicate actions, constraint violations, replanning time and operator interventions.

Explore the Other Programs