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

Summary

This paper proposes LOCUS-DT (Localization via Observation-Conditioned Uncertainty Scoring with Digital Twins), a framework that treats snapshot indoor localization as posterior inference over the transmitter location given measurements, a known room layout, and receiver pose. The method assumes the room layout and receiver pose are known, and uses a ray-tracing-based digital twin (DT) to generate synthetic multipath profiles for candidate transmitter locations, which are then compared against the measured channel profile to score each candidate location and obtain a posterior distribution.

The main contributions are threefold. First, the paper formulates single-snapshot indoor localization as posterior inference over the transmitter location, incorporating the layout implicitly through the ray-tracing DT to obtain multipath channel estimates for each candidate location. Second, it proposes a novel learned scoring function that compares the top K paths of the candidate and measured multipath profiles, with training performed over an ensemble of environments so the method can be evaluated on environments not included in the training set, incorporating errors in either the DT or channel estimation during training for robustness. Third, it validates the approach using a Sionna-based ray-tracing backend, benchmarking against Gaussian and Gaussian-mixture posterior models, demonstrating that explicit candidate-wise DT matching better captures the sharp, multimodal, and layout-dependent posterior structure induced by indoor multipath.

The localization setting is defined as follows: the transmitter is at an unknown position X t in R d (with d=2 for 2D localization), and the receiver is given (L, p) where L is a known room layout and p = (x r, phi r) is the receiver pose (position x r in R squared and orientation phi r in [-pi, pi)). At runtime, the receiver collects an observation y, typically an array of in-phase/quadrature (IQ) samples from the receiver antennas. The inference objective is to estimate the posterior distribution p(x t L, p, y), representing the relative likelihood of all potential transmitter locations given the observation, room layout, and receiver pose.

The DT multipath estimator is modeled abstractly as a function Psi(L, p, x t) = psi DT k, k = 1,..., K in S K, where each path is a tuple psi = (theta, tau, alpha, gamma) in S, with theta being the angle of arrival (AoA), tau the relative delay, alpha the complex path gain, and gamma = alpha squared / sigma squared the real-valued path SNR for receiver-noise variance sigma squared. The DT maps each candidate location x t to a predicted set of K paths, sorted from strongest to weakest. The abstraction is intentionally backend-agnostic, allowing any DT method to instantiate Psi as long as it produces the same ranked tuple representation.

The observation interface models the multipath estimator as a function Psi hat: C N x N rx -> S K, which produces the estimated multipath profile psi obs k, k = 1,..., K from the received IQ samples. The paper uses the iterative Levenberg–Marquardt (LM) algorithm for multipath estimation from array measurements.

The core modeling problem is constructing a logit function g omega(x t, L, p, y) = G omega(psi DT k(x t), psi obs k), where omega collects all trainable weights and biases of the neural network realizing the score function G omega. Assuming a prior distribution p 0(x t) (typically uniform over a region of interest), the posterior is given by p hat omega(x t L, p, y) = e g omega(x t,L,p,y) p 0(x t) / Z omega(y; L, p), where Z omega(y; L, p) is the normalizing constant obtained by integrating over all candidate locations.

The score function is trained using a sampled cross-entropy loss. Given a dataset D = L m, p m, C m, y m, x tm m=1 M, where each training sample includes a finite set of candidate transmitter locations C m = x bar tmj j=1 N m subset of R d, and the true transmitter location satisfies x tm in C m (with a unique candidate index c* m), the loss function is approximated as L(omega) = (1/M) sum m=1 M [-g mc* + log(sum j=1 N m e g mj)], where g mj = g omega(x bar tmj, L m, p m, y m). This is the candidate-set analogue of the sampled cross-entropy criterion in energy-based learning.

The score function G omega is realized as a transformation followed by a simple multilayer perceptron. In the simulations, K = 6 paths are retained from each DT and observation profile. Synchronization is unavailable and the system is narrowband, so delay is discarded. Each AoA theta is encoded as (cos theta, sin theta), while the SNR gamma is converted to dB and scaled to [0, 1] over a 60 dB range. The resulting 12 paths yield 3 x 12 = 36 features, fed into a two-layer neural network, providing an extremely compact, easily trainable scoring function with a limited number of parameters.

For experimental validation, the paper creates random synthetic environments with varying obstacles, noting the method is not site-specific and can be tested on completely random environments different from those used for training. Each random environment is a 100 m x 100 m region containing the outer room boundary and three randomly generated interior obstacles drawn from rectangular, T-shaped, and L-shaped primitives. The NVIDIA Sionna ray-tracing backend simulates the true channel and N m = 500 candidate channels. The carrier frequency is 12 GHz, signal bandwidth is 200 MHz, noise figure is 7 dB, synchronization window is 4 microseconds, and transmit power is 0 dBm. The transmitter uses one vertically polarized isotropic element, while the receiver uses a 1 x 8 vertically polarized half-wavelength array with a 3GPP TR 38.901 element pattern. The ray tracer returns up to 16 valid paths per channel; LOCUS-DT retains the K = 6 strongest paths from each DT candidate profile. For the online observation, the standard maximum-likelihood (SML) estimator uses a 128-point matched filter to estimate (theta, tau, alpha), derives gamma, and retains the K = 6 strongest scorer pairs (theta, gamma).

The training split spans 8000 distinct room layouts, each with 500 candidate transmitter locations and 10 observation realizations, for a total of 80,000 snapshots. The test split spans 2000 additional layouts with the same 500-candidate structure, yielding 20,000 snapshots. For the final benchmark comparison, two harder evaluation sets are reported: Eval Random uses 2000 held-out groups with 1000 feasible random candidates per group, while Eval Grid uses the corresponding held-out layouts but starts from a regular 31 x 31 raw grid and keeps 696 layout-feasible candidate positions after pruning invalid points.

The maximum-likelihood observation estimator uses frequency- and spatial-domain basis functions b F(tau) = e-j2pi f 0 tau * m m=0 M F - 1 and b R(theta) = e-j pi sin theta * m m=0 M R - 1, where M F is the number of subcarriers, f 0 is the subcarrier spacing, and M R is the number of half-wavelength-spaced receive-ULA elements. The frequency-space channel response over K paths is s(Psi) = sum k=1 K b R(theta k) tensor b F(tau k) alpha k. Assuming equal-power pilots across subcarriers, the channel observation is modeled as y m = s(Psi) + w m, where w m is additive white Gaussian noise with variance sigma squared. The maximum-likelihood estimate is Psi hat = arg max Psi p(y m Psi).

The paper compares against three benchmarks: Gauss-Cart (Gaussian posterior with transmitter location in Cartesian coordinates, with mean mu and covariance Q as functions of the same full inputs), Gauss-Polar (identical estimator but with posterior in polar coordinates), and GMM-Cart (mixture with k = 3 Gaussian components). All trainable models are optimized with AdamW for 2000 epochs with weight decay 10-4 and hidden-layer dimensions (64, 16). LOCUS-DT uses pairwise feature construction whereas Gaussian benchmarks use candidate-independent conditioning on receiver pose and IQ observation. Learning rates are 5 x 10-3 for LOCUS-DT and 2 x 10-3 for Gaussian baselines. Training runs on a single NVIDIA A100 GPU, requiring approximately 2.5 GPU-hours for LOCUS-DT, 1.5 GPU-hours for Gauss-Cart, 1.5 GPU-hours for Gauss-Polar, and 2.0 GPU-hours for GMM-Cart, totaling approximately 7.5 GPU-hours.

Besides the training cross-entropy, the paper reports the adjusted loss L = (1/M) sum m=1 M [-log p hat mc* - log N m], which measures gain relative to uniform random guessing over the candidate set, and derived quantities G = e-L and R = -L / log N cand, where N cand is the split-specific candidate count. Lower L and higher G and R indicate better performance.

Qualitative results show LOCUS-DT on three distinct random layouts, with each row pairing a Sionna room with posterior heatmaps for six TX/RX configurations. LOCUS-DT yields the most localized and geometry-consistent posteriors, with high-probability regions aligned with the true transmitter and wall- and reflection-induced structure. Gauss-Cart and Gauss-Polar produce smooth single-mode fields that miss sharp spatial discontinuities and multimodal ambiguity. GMM-Cart captures limited multimodality, but its peaks remain broader and less aligned because it lacks candidate-wise DT matching. Quantitative results in Table I confirm that LOCUS-DT outperforms all Gaussian baselines on both evaluation sets, with LOCUS-DT achieving adjusted loss values of-2.7224 (Eval Grid) and-3.3798 (Eval Random), compared to Gauss-Cart (-0.0594, -0.1190), Gauss-Polar (-0.3030, -0.2650), and GMM-Cart (-0.3778, -0.5078), with correspondingly higher G and R values.

The paper concludes that LOCUS-DT captures reflection-induced multimodal ambiguity, outperforms Gaussian and Gaussian-mixture baselines, and generalizes to unseen layouts. Future work will evaluate greater DT mismatch and measured rooms reconstructed using SLAM.

Based on the paper, here are the specific improvements I can make to AI systems and what the improved system can do:

1. Physics-informed posterior inference for localization

• Replace single-point regression outputs with a learned scoring function that compares ray-tracing digital twin (DT) predictions against observed multipath profiles

• Train the scoring function over an ensemble of environments (8000 layouts) so it generalizes to unseen room geometries without retraining

• Use a candidate-set cross-entropy loss with Monte Carlo partition function estimation to handle intractable normalization

2. Robust multimodal uncertainty representation

• Encode each path as (cos θ, sin θ, SNR) features, discarding delay for narrowband synchronization-free operation

• Retain top K=6 paths from both DT and observation, producing a 36-dimensional feature vector fed into a two-layer MLP (64, 16 hidden units)

• Incorporate DT and channel estimation errors during training to maintain robustness against model mismatch

3. Layout-conditioned candidate scoring

• Generate 500 candidate locations per scene using Sionna ray tracing at 12 GHz, 200 MHz bandwidth

• Compare DT-predicted multipath signatures against maximum-likelihood estimates from 1×8 receive array snapshots

• Output a full posterior heatmap over the 2D space rather than a single point estimate

1. Achieve significantly better localization accuracy than Gaussian baselines

• On held-out random layouts with 1000 candidates: adjusted loss L = −3.3798 vs −0.1190 (Gauss-Cart), −0.2650 (Gauss-Polar), −0.5078 (GMM-Cart)

• Achieve 48.93% relative gain over uniform guessing (R metric) vs 1.72% for the best Gaussian baseline

• On grid-based candidates (696 positions): L = −2.7224, R = 41.59%, outperforming all baselines by over 35 percentage points

2. Capture sharp, multimodal, layout-dependent posterior structures

• Produce posterior heatmaps that align high-probability regions with true transmitter locations and reflection-induced ambiguity patterns

• Resolve spatial discontinuities and multiple peaks that Gaussian or Gaussian-mixture models smooth over

• Maintain geometry-consistent uncertainty even when the true location is ambiguous due to heavy blockage or multipath

3. Generalize to unseen environments without site-specific retraining

• Train once on 8000 random layouts, then evaluate on 2000 completely different held-out layouts

• Handle arbitrary room shapes (rectangular, T-shaped, L-shaped obstacles) and any receiver pose

• Operate with only a known layout and receiver pose, requiring no fingerprinting or prior measurements in the target environment

4. Provide actionable uncertainty for downstream tasks

• Output a normalized posterior distribution over all feasible transmitter locations, enabling robotic navigation, search-and-rescue, and multi-hypothesis tracking

• Support decision-making under ambiguity by exposing multimodal hypotheses rather than forcing a single estimate

• Enable integration with planning algorithms that require probabilistic spatial beliefs (e.g., reinforcement learning for indoor navigation)

Accurate indoor localization is essential for emerging applications in robotic navigation and search and rescue. While classical methods typically focus on single-point estimates, complex indoor environments with heavy blockage and multipath propagation often lead to multimodal likelihood surfaces where a single estimate is insufficient. This paper proposes LOCUS-DT (Localization via Observation-Conditioned Uncertainty Scoring with Digital Twins), a framework that treats snapshot localization as posterior inference over the transmitter location. By leveraging a ray-tracing-based digital twin (DT) of the known environment, LOCUS-DT generates synthetic multipath profiles for candidate locations and compares them against the measured channel profile. Central to our approach is a novel learned scoring function designed to compare a fixed number of dominant specular paths, providing robustness against errors in both the DT environment model and the physical channel estimation. Importantly, LOCUS-DT is trained over an ensemble of environments to ensure generalization to unseen layouts. We evaluate the system using a Sionna-based ray-tracing backend, demonstrating that LOCUS-DT captures the sharp, multimodal posterior structures inherent in indoor settings more accurately than standard Gaussian or Gaussian-mixture benchmarks.