⚡ electrical engineering

LOCUS-DT: Localization via Observation-Conditioned Uncertainty Scoring with Digital Twins

This paper introduces LOCUS-DT, a framework that leverages ray-tracing-based digital twins and a novel learned scoring function to perform robust, multimodal posterior inference for accurate indoor localization in complex environments with heavy blockage and multipath propagation.

Original authors: Haozhe Lei, Roberto Bomfin, Marwa Chafii, Sundeep Rangan

Published 2026-08-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haozhe Lei, Roberto Bomfin, Marwa Chafii, Sundeep Rangan

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to find a friend in a giant, empty warehouse, but you can't see them. You only have a walkie-talkie. If you shout "Hello!" and they shout back, you might guess where they are based on how loud their voice is. But what if the warehouse is full of shiny metal walls, giant pillars, and weird corners? Your friend's voice might bounce off a wall, hit a pillar, and come back to you from a completely different direction than where they actually are. In the real world, radio waves behave exactly like that voice. They bounce, scatter, and create a confusing mess of echoes. This is the challenge of "wireless localization": figuring out exactly where a device is just by listening to its radio signals.

For a long time, scientists tried to solve this by guessing a single "best" spot, like drawing one X on a map. But in complex indoor spaces, that single guess is often wrong because the radio signals are so tricky. Instead, modern researchers are starting to think of localization as a game of probability: "Here is a map of all the places the device could be, and here is how likely it is to be in each spot." To do this, they use "Digital Twins"—super-smart computer simulations that act like a virtual copy of the real room, predicting exactly how radio waves should bounce around if the device were in any specific spot. The big question is: how do we compare the real, messy radio signals we hear with the perfect, clean signals the computer predicts, especially when the computer isn't 100% perfect and the real world is full of surprises?

This paper introduces a new system called LOCUS-DT (Localization via Observation-Conditioned Uncertainty Scoring with Digital Twins) to solve that exact problem. Think of LOCUS-DT as a super-smart detective who doesn't just look for one clue, but compares the entire story of the radio echoes.

Here is how the detective works: First, the system knows the layout of the room (where the walls and furniture are) and where the listening device is standing. It then creates a "library" of possibilities. For every single spot the transmitter could be hiding, the system uses its Digital Twin (a ray-tracing simulation) to predict what the radio signal should look like if the transmitter were actually there. It calculates the top few strongest "echoes" (paths) that would bounce off walls and arrive at the listener.

Next, the system listens to the real signal from the transmitter and extracts its own set of top echoes. Now comes the magic trick: instead of just checking if the distances match, LOCUS-DT uses a learned scoring function. Imagine this as a highly trained judge who looks at the "top 6" echoes from the real world and the "top 6" echoes from the computer simulation for a specific spot. The judge compares them side-by-side. If the real echoes look very similar to the simulated echoes for that spot, the judge gives it a high score. If they look nothing alike, the score is low.

The paper shows that this method is much better than older ways of guessing. Older methods often assume the answer is a simple, smooth blob (like a Gaussian distribution), which works okay in open fields but fails miserably in rooms with walls because real radio signals create sharp, confusing "ghost" locations. LOCUS-DT, however, is trained on thousands of different fake rooms with random obstacles. This training teaches the system to be robust against errors. Even if the computer simulation isn't perfect or the real signal is a bit noisy, the system learns to ignore the mismatches and focus on the patterns that truly matter.

The results, tested in detailed computer simulations using a tool called Sionna, show that LOCUS-DT creates a much more accurate "heat map" of where the transmitter is. Instead of a blurry, single spot, it produces a sharp, multi-peaked map that correctly identifies the true location while also showing other possible "ghost" locations caused by reflections. It outperforms standard methods that try to fit simple shapes to the data, proving that by treating localization as a matching game between real and simulated echoes, we can get a much clearer picture of where things are, even in the most confusing indoor environments. The authors emphasize that this was demonstrated through simulations, suggesting that this approach could be a powerful tool for future robots and search-and-rescue teams navigating complex buildings.

Technical Summary: LOCUS-DT

Problem Statement
Accurate indoor localization is critical for applications such as robotic navigation and search and rescue. However, classical localization methods typically provide a single-point estimate of a transmitter's (TX) location. In complex indoor environments characterized by heavy blockage and multipath propagation, the likelihood surface often becomes multimodal, rendering single-point estimates insufficient. Furthermore, many data-driven methods are tightly coupled to specific training environments, limiting their robustness when layout-dependent propagation structures change. The core challenge addressed is estimating the full posterior distribution of the TX location given a single measurement snapshot, a known room layout, and the receiver (RX) pose, rather than merely finding a point estimate.

Methodology: LOCUS-DT
The paper proposes LOCUS-DT (Localization via Observation-Conditioned Uncertainty Scoring with Digital Twins), a framework that treats snapshot localization as posterior inference over the TX location. The methodology operates under the assumption that the room layout ( $L$ ) and RX pose ( $p$ ) are known.

Digital Twin (DT) Generation: A ray-tracing-based digital twin is used to generate synthetic multipath profiles for a set of candidate TX locations. For any candidate location $x_t$ , the DT predicts a set of $K$ dominant specular paths, represented as tuples of angle of arrival (AoA), relative delay, complex path gain, and path SNR.
Observation Extraction: From the received signal (IQ samples), an online estimator (using an iterative Levenberg–Marquardt algorithm) extracts a measured multipath profile, also consisting of the top $K$ paths.
Learned Scoring Function: A novel neural network-based scoring function, $g_\omega$ , is trained to compare the DT-predicted multipath signature of a candidate location against the observed multipath signature. This function outputs a compatibility score.
Posterior Inference: The posterior distribution over candidate locations is constructed by normalizing the exponential of these scores (Boltzmann distribution). The normalization constant is approximated using a Monte Carlo estimate over the finite set of candidate locations during training.
Training Strategy: The model is trained using a sampled cross-entropy loss over an ensemble of diverse synthetic environments. This approach ensures the method generalizes to unseen layouts. The training data consists of candidate sets where the true TX location is one of the candidates, allowing the network to learn to distinguish the correct location based on multipath matching.

Key Contributions

DT-Conditioned Posterior Formulation: The work formulates single-snapshot indoor localization as posterior inference where the room layout is implicitly incorporated via a ray-tracing digital twin. This allows for the comparison of candidate-specific multipath estimates against measured profiles.
Generalizable Learned Scoring Function: A new scoring function is introduced that compares the top $K$ paths of candidate and observed profiles. Crucially, because training occurs over an ensemble of environments, the method is not site-specific and can be evaluated on layouts not present in the training set. It also incorporates robustness against errors in both the DT model and channel estimation.
Validation with Realistic Ray Tracing: The system is instantiated using a Sionna-based ray-tracing backend and benchmarked against standard Gaussian and Gaussian-mixture posterior models.

Experimental Results
The system was evaluated using a Sionna-based backend with 12 GHz carrier frequency and 200 MHz bandwidth. The training set included 8,000 distinct room layouts, while the test set included 2,000 unseen layouts.

Qualitative Performance: Visualizations of the posterior heatmaps demonstrate that LOCUS-DT captures sharp, multimodal structures inherent to indoor multipath (e.g., reflections off walls). In contrast, Gaussian benchmarks (Cartesian and Polar) produce smooth, single-mode fields that fail to capture spatial discontinuities, and Gaussian Mixture Models (GMM) capture limited multimodality but with broader, less aligned peaks.
Quantitative Performance: On evaluation sets with 696 and 1,000 candidates, LOCUS-DT significantly outperformed all Gaussian and GMM baselines. It achieved a lower adjusted loss ( $L$ ) and higher gain ( $G$ ) and relative performance ( $R$ ) metrics. For instance, on the "Eval Grid" set, LOCUS-DT achieved an adjusted loss of -2.72 compared to -0.06 for the best Gaussian baseline, and a relative performance of 41.59% versus 0.91%.

Significance and Claims
The paper claims that LOCUS-DT successfully addresses the ambiguity of indoor localization by explicitly modeling the layout-dependent, multimodal nature of the posterior distribution. By leveraging a digital twin to generate candidate hypotheses and a learned scoring function to match them against observations, the framework captures the "sharp, multimodal posterior structures" that standard Gaussian assumptions miss. The authors emphasize that their approach moves beyond generic class logits to candidate-wise compatibility scores, providing a more robust and generalizable solution for uncertainty quantification in indoor RF localization. Future work is noted to involve evaluating the system under greater DT mismatch and using real-world rooms reconstructed via SLAM.

Technical Summary: LOCUS-DT

More like this