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.

Verification Human Oversight and Emerging Connectivity

This research program is conducted by a team of Ph.D. researchers with expertise in human oversight and emerging connectivity.

Develop evidence that decisions and actions remain within stated limits. Address human workload, communication failures and the boundaries of any formal or empirical assurance.

Runtime Assurance Human Oversight and Physical Outcome Verification

Check proposed actions and observed responses against permitted operating conditions. Investigate independent monitors, escalation policies, operator interfaces and application specific fallback states. Assess both unnecessary interventions and failures to intervene, including realistic operator response time. For mobile platforms, assess stopping, controlled landing, restricted motion and supervised handover as application specific responses; no single fallback is appropriate to every machine.

Research questions

When should the system act, abstain or request approval? Can a monitor detect an ineffective action even when equipment reports successful execution?

Proposed Demonstration

Introduce failed actuators, conflicting readings and unexpected responses into a bounded control rig with operator oversight.

Evaluation

Detection coverage, intervention latency, false intervention rate, operator workload and successful transition to the defined fallback state.

Bounded Formal Verification and Foundation Model Action Assurance

Translate open ended model proposals into checkable commands, state transitions and permission checks. Investigate formal properties of restricted action interfaces, including ordering constraints and forbidden states. Record model assumptions and separate verified software properties from unproven physical behavior.

Research questions

Which useful properties can be specified precisely? How does incomplete state information change the meaning of a verified precondition?

Proposed demonstration

Test a restricted action interface against erroneous and adversarial proposals, including attempts to bypass required sequencing.

Evaluation

Property coverage, invalid actions rejected, valid actions wrongly rejected, verification overhead and explicit limits of the guarantees.

Wireless Sensing Integrated Sensing and Communication and Communication Aware Physical AI

Study how interference, packet loss, latency, clock uncertainty and limited bandwidth affect sensing and supervisory control. Compare appropriate existing wireless links and examine integrated sensing concepts as experimental extensions. Future connectivity capabilities are research variables rather than deployment dependencies. Investigate integrated sensing and communication using suitable programmable radio or UWB testbeds. Distinguish shared hardware from signals that jointly carry data and support sensing. Study resource allocation between communication and sensing, multipath, interference and clock synchronization. Treat future 5G-Advanced or 6G features as experimental options whose actual availability must be established.

Research questions

Which decisions require fresh local measurements? How should stale data change confidence, operating limits and fallback behavior? What tradeoffs arise between sensing error, detection coverage, packet reliability and latency under realistic industrial conditions? Set application specific targets rather than presuming subcentimeter accuracy or deterministic wireless timing.

Proposed demonstration

Emulate delay, jitter, packet loss and outages across sensor and gateway links on a bounded control rig. Add a contained RF or recorded signal experiment comparing separate and shared sensing and communication resources under varying traffic and multipath.

Evaluation

End to end latency percentiles, measurement age, decision degradation, limit violations, availability and recovery after connectivity returns. Report positioning error distributions, presence detection, sensing resolution, missed deadlines and confidence intervals by condition.

Human Factors Explainability and Operator Decision Support

Design explanations that connect recommendations to relevant observations, uncertainty, constraints and feasible alternatives. Study alarm fatigue, automation bias and handovers between advisory and automated modes. Evaluate comprehension through realistic operational tasks rather than satisfaction scores alone.

Research questions

Can operators detect a plausible but incorrect recommendation? Which explanation details improve decisions without increasing response time excessively?

Proposed demonstration

Compare interfaces in simulated maintenance and control incidents, including incorrect AI advice and ambiguous evidence.

Evaluation

Correct interventions, response time, comprehension, workload, inappropriate acceptance of AI advice and quality of handover decisions.

Standards Aligned Assurance Evidence and Regulatory Readiness

Investigate how research evidence can support later safety, security and conformity reviews. Link hazards, requirements, controls, tests, measured results and limitations. Select applicable frameworks and editions with qualified practitioners; candidate references include IEC 62443, ISO 10218, ISO 3691-4 and IEC 61508 where relevant to the application. Maintain an applicability register for the intended market, including the EU Machinery Regulation where relevant, with verified effective dates and assigned owners. Track evidence as models, data, software and operating conditions change. Research evidence alone does not establish certification, compliance or functional safety.

Research questions

Which requirements in existing standards remain difficult to satisfy when behavior is learned rather than explicitly programmed? How can evidence remain valid and traceable after a model update without repeating the complete assessment?

Proposed demonstration

Construct an evidence package for one bounded application, such as the restricted equipment control gateway of Topic 14, linking hazards, controls, test scenarios, results and residual limitations. Introduce a model update and assess which evidence must be regenerated.

Evaluation

Requirement coverage, traceability completeness, unresolved gaps, effort to update evidence after a change, reviewer assessment by qualified safety and security practitioners and reuse of the evidence structure across applications.

Hybrid Terrestrial and Non Terrestrial Connectivity for Remote Assets

Investigate telemetry and supervisory coordination across available terrestrial links and satellite or other non terrestrial links for remote industrial assets. Begin with measured or emulated link profiles. Study store and forward delivery, state freshness, handover, cost and loss of connectivity. Keep protective functions and time critical control local; continuous telemetry and seamless handover are outcomes to test, not assumptions.

Research questions

Which functions can continue during intermittent satellite fallback? How should delayed or uncertain state restrict remote command authority and the permitted local operating envelope?

Proposed demonstration

Emulate terrestrial to satellite transitions on a remote asset or mobile platform simulator. Inject delay, jitter, outages, reordering and duplicate messages; compare fixed policies with freshness aware autonomy limits before any field pilot.

Evaluation

Telemetry age, delivery completeness, handover interruption, stale command rejection, retained local functions, recovery time, bandwidth cost and operating limit violations.

Explore the Other Programs