Probabilistic Inference | Haozhe Lei

MC-CLE: Full-Posterior Wireless Localization

Thu, 01 Jan 2026 00:00:00 +0000

I lead MC-CLE, a probabilistic localization method that asks how compatible each candidate transmitter location is with a sparse angle-and-strength measurement. Instead of forcing ambiguous RF evidence into one coordinate, MC-CLE learns a spatial likelihood field that retains directional structure and multiple plausible hypotheses.

From physical measurement to spatial likelihood

The model conditions candidate-wise scoring on receiver geometry and the measured channel signature. Normalizing the scores over candidate locations yields a full posterior that can be calibrated and fused across measurements.

Journal schematic. Scene geometry, receiver pose, and the RF channel signature condition a learned candidate scorer, producing a full spatial posterior rather than a single estimate.

Conference evidence. Across varied transmitter-receiver geometries, MC-CLE preserves directional and multimodal posterior structure that Gaussian point-error models smooth away or constrain incorrectly.

The peer-reviewed introduces likelihood-based full-posterior localization. The first- and corresponding-author journal extension, , is under review at IEEE Transactions on Vehicular Technology (TVT).

Relationship to LOCUS-DT

MC-CLE and address the same full-posterior localization objective through complementary model classes. MC-CLE learns a measurement-conditioned candidate likelihood, while LOCUS-DT explicitly compares observations with ray-traced candidate signatures. Both provide calibrated spatial belief as an interface to downstream localization and autonomous sensing.

LOCUS-DT: Digital-Twin Posterior Localization

Thu, 01 Jan 2026 00:00:00 +0000

I lead LOCUS-DT, a wireless-localization framework that compares an observed multipath snapshot with candidate-specific signatures generated by a ray-tracing wireless digital twin. The learned compatibility scores form a posterior over transmitter location, including multiple plausible modes when the measurement is incomplete or the model is imperfect.

Observation-conditioned digital-twin inference

For every candidate location, the digital twin supplies the multipath structure that the receiver would expect. LOCUS-DT scores the agreement between those simulated hypotheses and the measured path peaks, then normalizes the scores into a calibrated spatial belief.

Journal schematic. LOCUS-DT matches measured multipath features with candidate-specific ray-traced signatures and converts their learned compatibility into a transmitter-location posterior.

Conference evidence. Experiments across three unseen indoor layouts recover structured, often multimodal posteriors that expose blockage and multipath instead of hiding them behind one coordinate.

The peer-reviewed establishes LOCUS-DT. The first- and corresponding-author journal extension, , studies site generalization and explicit digital-twin mismatch; a revision is in preparation for resubmission to IEEE Transactions on Wireless Communications (TWC).

Relationship to MC-CLE

LOCUS-DT and share a full-posterior localization objective but use complementary models. LOCUS-DT explicitly evaluates observed multipath against ray-traced candidate signatures; MC-CLE learns a measurement-conditioned candidate likelihood. Their common output is calibrated spatial belief that can support localization, multi-view fusion, and active sensing.

MAGNETAR: Joint RF Pose Inference

Thu, 17 Sep 2026 00:00:00 +0000

I lead MAGNETAR from probabilistic formulation and neural architecture design through real-to-sim calibration, MobiFR3 integration, experiment design, and evaluation. The method estimates a joint posterior over a transmitter’s planar position and heading from one asynchronous RF multipath snapshot, rather than forcing an ambiguous observation into a single pose estimate.

Your browser does not support embedded MP4 video.

The three-minute overview shows why position and heading must be inferred together, how MAGNETAR scores pose hypotheses, and how it is evaluated with robotic measurements.

Joint pose inference. A heading-conditioned spatial scorer preserves competing position and orientation hypotheses rather than collapsing an ambiguous RF snapshot into one estimate.

Physical validation and fusion. MobiFR3 couples 10-GHz measurements with robot poses; evidence from multiple viewpoints is fused to resolve ambiguous transmitter poses.

Method

Represents each observation by angle-of-arrival and signal-to-noise-ratio estimates, together with the known room layout and receiver pose.
Uses a heading-conditioned shared 2D U-Net to score and jointly normalize a discretized position-heading grid.
Trains primarily in real-to-sim-calibrated 10-GHz simulation, with measured antenna patterns and RF-chain/noise effects, then augments training with 1,800 measured records.
Fuses joint posteriors across observations so that evidence with incompatible heading hypotheses does not reinforce the wrong position.

Physical evaluation

The experimental study uses , the robotic RF system that I lead. On 24,000 held-out measurements, the heading-conditioned scorer achieved the lowest real-data negative log-likelihood and the highest joint hit rate among the tested models. Student collaborators whom I mentor contributed to RF operation and data acquisition; I led the algorithm, calibration, RF-robot integration, experimental design, and evaluation.

Paper: — first and corresponding author; IEEE ICRA 2027, under review.

Beyond Point Estimates: Likelihood-Based Full-Posterior Wireless Localization

Thu, 01 Jan 2026 00:00:00 +0000

First and corresponding author

MC-CLE scores candidate transmitter states rather than regressing directly to one coordinate. The resulting posterior retains spatial ambiguity created by blockage, antenna directivity, and multipath, and provides a probabilistic interface for multi-view fusion and downstream decisions.

The continuing journal study, , is under review at IEEE Transactions on Vehicular Technology.