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 Identification Sensing and State Estimation

Establish dependable identity, location, condition and process context despite imperfect observations. Selected work can strengthen the interface between the Identification Engine and Sensing Engine.
This research program is conducted by a team of Ph.D. researchers with expertise in identification and sensing technologies.

Multimodal Industrial Identification Localization and State Estimation

Combine RFID, UWB, BLE, vision and equipment signals to associate observations with physical objects. Investigate timestamp alignment, coordinate transformations, probabilistic data association and uncertainty propagation. Address occlusion, metal interference, tag collisions and handoffs between indoor and outdoor areas. Include object and human tracking around robots, drone observation alignment and vehicle localization where those platforms supply measurements.

Research questions

When should conflicting observations be fused or rejected? How can identity be recovered after a long observation gap without assigning history to the wrong asset?

Proposed Demonstration

Track materials from receiving through storage to a production station; introduce missed reads, overlapping movements and camera occlusion. Compare fusion with each individual modality.

Evaluation

Identity switches, localization error distribution, event precision and recall, recovery time and confidence calibration, including performance by zone and interference condition.

AIoT Based Inventory Asset Movement and Process Understanding

Infer receipts, transfers, consumption and work in progress transitions from incomplete event streams. Combine movement evidence with work orders, expected routes and stock balances. Study duplicate suppression, delayed events and reconciliation between observed physical state and enterprise records.

Research questions

Can the system distinguish a temporary observation gap from a missing item? How should later evidence revise an earlier inventory estimate while preserving its audit history?

Proposed demonstration

Use a stockroom and simulated production area with partial picks, returns and unrecorded transfers; compare results with independently counted inventory.

Evaluation

Inventory accuracy, transfer detection, discrepancy resolution time, manual reconciliation effort and errors caused by delayed data.

Adaptive Sensor Placement Sampling and Active Perception

Select sensor locations, sampling rates and inspection viewpoints according to decision value, power and installation constraints. Evaluate information gain, physical observability and redundancy under sensor failure. Study when a system should request another measurement instead of acting on incomplete evidence.

Research questions

Which measurements reduce uncertainty enough to change a maintenance or control decision? Can adaptive sampling preserve detection of short events while reducing power and bandwidth?

Proposed demonstration

Compare fixed and adaptive sampling on a motor or environmental rig; remove selected sensors and vary event duration.

Evaluation

Detection coverage, missed short events, uncertainty reduction, battery or compute consumption and installation cost for equivalent decision quality.

Industrial Sensor Reliability and Data Quality Intelligence

Detect drift, calibration errors, unit mismatches, clock errors, communication gaps and sensor substitution. Combine redundant observations with physical relationships to separate sensor faults from equipment faults. Propagate measurement quality into downstream predictions and permission to act.

Research questions

How early can gradual drift be distinguished from real process change? What level of uncertainty should trigger recalibration, alternative sensing or suspended automation?

Proposed demonstration

Inject bias, noise, stale values and timing errors into an instrumented pump or thermal rig with an independent reference sensor.

Evaluation

Fault detection delay, false alarms, isolation accuracy, calibration recovery and degradation in downstream decisions relative to clean data.

Physical Traceability Product Genealogy and Chain of Custody

Link material batches, components, tools, process conditions and custody transfers across splitting, merging, rework and repackaging. Investigate event graphs and reconciliation of inconsistent identifiers across enterprise systems. Keep physical identity evidence distinct from asserted records.

Research questions

How can uncertain links be represented without creating false certainty about provenance? What evidence is needed to localize an affected batch after mixed production?

Proposed demonstration

Trace sample batches through assembly, rework and shipment, including a deliberately incorrect tag association and missing handoff.

Evaluation

Genealogy completeness, incorrect associations, recall scope accuracy, reconstruction time and visibility of unresolved custody gaps.

RFID and RTLS World State Grounding for Embodied AI

Maintain a persistent world state of asset identity, location, custody, condition and task status that robot policies, vision language action models and industrial agents can query before and during action. Investigate how identification and positioning evidence is combined with perception to resolve which physical object an instruction refers to, whether an object has moved since it was last observed and whether its recorded state is still current. Represent uncertainty and observation age explicitly so that a planner can distinguish confirmed state from assumed state. Study how the world state is updated by the outcomes of robotic and equipment actions. Treat tag identifiers as identity evidence, not authentication. Keep observation source, timestamp and confidence with each state update.

Research questions

Does identification and positioning context reduce wrong object selection and stale assumptions compared with perception alone? How old or uncertain may a world state entry become before an agent must reobserve the object rather than act on the stored record?

Proposed demonstration

Instruct a robot arm or mobile platform to retrieve and deliver tagged items in an RTLS instrumented area containing visually similar objects, relocated items and deliberately outdated records. Compare a perception only policy, a policy with world state context and a policy that also enforces observation age limits.

Evaluation

Correct object selection, task success, actions taken on stale state, reobservation requests, world state accuracy after task completion and the latency added by world state queries.

Battery Free RFID Backscatter Sensing and Energy Harvesting Intelligence

Use passive and energy harvesting RFID tags as sensing elements rather than identifiers alone. Investigate inference of temperature, strain, moisture, liquid level, vibration, touch and movement from dedicated sensor tags and from changes in tag phase, signal strength and read patterns. Address sensitivity to reader geometry, multipath, metal surfaces, nearby tags and environmental change. Target assets where wired or battery powered sensors are impractical, including sealed containers, rotating equipment, embedded structures and high density storage. Extend selected experiments to nodes powered by thermal gradients or vibration, event triggered inference and intermittent computing. Battery free does not mean energy free. Distinguish physical sensing by dedicated tags from indirect inference using radio measurements; each quantity requires its own feasibility study.

Research questions

Which physical quantities can be estimated reliably from backscatter behavior without dedicated sensor hardware, and which require sensor tags? How can the system separate a genuine change in the measured quantity from a change in reader position, interference or surrounding objects? How can checkpointing, sequence identifiers and restart reconciliation preserve valid state across interrupted power without inventing continuous observations?

Proposed demonstration

Instrument a liquid container, a strain specimen and a rotating fixture with standard and sensor enabled tags alongside independent reference sensors. Vary reader distance, antenna orientation, nearby metal and tag density. Begin with one measurable quantity; add the other fixtures only after feasibility is established. Interrupt harvested power and assess checkpoint recovery against continuously powered references.

Evaluation

Measurement error against reference sensors, detection sensitivity, false change detection under interference, read reliability, calibration effort per installation and cost per sensing point relative to conventional sensors. Also measure energy availability, cold start time, duty cycle, missed events and state corruption after power loss.

Perception in Degraded Industrial Environments Using Radar Thermal and Event Based Sensing

Evaluate millimetre wave radar, thermal imaging and event based cameras for industrial settings in which conventional cameras degrade, including dust, steam, smoke, darkness, glare, vibration and rapid motion. Investigate occupancy and presence detection near machinery, obstacle detection for mobile platforms and inspection of heated or fast moving equipment. Study fusion with conventional cameras and depth sensors, and the detection of conditions under which each modality becomes unreliable. Compare modalities separately under each condition; none is assumed to see through every obscurant. Experimental detection does not replace independently qualified protective sensing.

Research questions

Under which environmental conditions does each modality provide decisive evidence that visible cameras cannot? Can a perception system recognize its own degradation early enough to reduce speed, request another modality or suspend an automated action?

Proposed demonstration

Operate a sensor suite in a controlled test area with generated fog or particulates, variable lighting, heated surfaces and moving personnel mannequins or targets. Compare single modality and fused perception for presence detection and obstacle recognition. Use contained, nonhazardous obscurants and surrogate targets, with no personnel exposed to experimental machine motion.

Evaluation

Detection coverage by environmental condition, missed presence events, false detections, degradation detection delay, latency, compute requirements and hardware cost for equivalent detection performance.

Our eight research programs address complementary aspects of Physical AI and AIoT and have been performed by a team of experienced Ph.D. researchers. Explore the other programs below.

Explore the Other Programs