Indoor radio positioning with ultra-wideband (UWB) radios, which time short pulses to fractions of a nanosecond, works by measuring time-difference-of-arrival (TDOA): how much later a tag hears one fixed beacon ("anchor") than another. Every nanosecond of unexplained timing error becomes about 30 cm of position error. Much of that error is not random. Each anchor carries its own slowly changing timing bias, from antenna delay, imperfect clock-offset correction, and delays that depend on received signal power. Today this bias is either calibrated away by hand for each station or left in the measurements. We propose estimating it from the flight data itself, with no motion capture and no per-station calibration.
The key observation is structural. Any error that belongs to an anchor rather than to a link enters a pairwise difference as $b_i - b_j$. Across all anchor pairs, these errors therefore form a skew-symmetric matrix of rank exactly two. Because the biases also drift slowly, the anchor-by-time bias table is low-rank in time as well. Estimating this table jointly with the tag's trajectory gives a large, redundant, nearly rank-deficient least-squares problem. Truncated singular value decomposition (truncated SVD) is the natural tool for such problems: it solves the system using only its well-determined directions and discards the rest. It also tells us exactly which combinations of anchor biases the trajectory so far can and cannot reveal.
Core Claim
Two things are claimed, and both can be tested on public data.
The first claim is empirical. On the UTIL UWB TDOA dataset (UTIL), a substantial share of the residual between measured TDOA and the TDOA implied by motion-capture ground truth is explained by an anchor-level term that is skew rank-2 at each instant and smooth over time. If this is false, nothing downstream matters, and the first experiment is designed to find that out cheaply.
The second claim is methodological. A sliding-window estimator can solve jointly for tag positions and a low-dimensional anchor-bias field, regularized by truncating the singular spectrum of the joint Jacobian (the matrix of how measurements change with each unknown). Such an estimator should recover most of the gain an oracle calibration would give, without that oracle. The singular spectrum also yields an observability map: for each window, it states which bias combinations were identifiable.
Concretely, for anchors $i,j$ at time $t$ with tag position $p(t)$ and anchor positions $s_i$, the residual model is
$$d_{ij}(t) - \big(\|p(t)-s_i\| - \|p(t)-s_j\|\big) = g_i(t) - g_j(t) + o_{ij}(t) + n_{ij}(t), \qquad g_i(t) = \sum_{k=1}^{K} c_{ik}\,\phi_k(t).$$
Here $d_{ij}$ is the measured range difference, $g_i$ is anchor $i$'s bias expressed in a small smooth temporal basis $\phi_k$ (constant, linear drift, a few slow terms), $o_{ij}$ is a sparse, nonnegative link-specific non-line-of-sight (NLOS) delay, and $n_{ij}$ is noise.
Why This Is Different From What Exists
The rank-2 fact comes from Ranking and synchronization from pairwise measurements via SVD. That paper recovers static offsets $r$ from noisy pairwise differences by projecting onto the top two singular vectors of $H = re^T - er^T$, and it names clock synchronization as a motivating application. We reuse that structure but change the problem in three ways. The offsets drift in time. They are entangled with an unknown moving position rather than observed directly. And the all-ones direction $e$ appears as an exact gauge freedom of the joint system, a direction no measurement can pin down, rather than as a nuisance to project out.
Translation Synchronization via Truncated Least Squares (TranSync) handles heavy biased outliers by iteratively zeroing high-residual edges in a 1D graph problem. It truncates measurements, not singular values. We borrow its residual gating for the sparse NLOS term $o_{ij}$, but it cannot tell us which bias directions are identifiable.
The case for estimating only the well-conditioned directions of a redundant parameterization comes from numerical approximation. Frames and Numerical Approximation and Stable and Accurate Least Squares RBF Approximations on Bounded Domains show that discarding small singular values makes redundant, ill-conditioned fits accurate. Fourier extensions from fixed data adds a caution: the truncation level trades stability against resolution. Here that means a truncation set too high would throw away real, slowly excited bias components. We treat this as a known risk to measure, not something to assume away.
The physical motivation comes from the UWB hardware papers. Decawave UWB Clock Drift Correction and Power Self-Calibration shows that power-dependent delays corrupt drift estimates, and reports an uncompensated offset of 0.3 m in one example. They remove it with per-station calibration curves; we aim to remove the anchor-level part of it online. Asynchronous Hyperbolic UWB Localization cancels linear clock drift algebraically and still reports degraded vertical accuracy from near-planar anchors. That is exactly the geometry in which anchor bias and vertical tag position become hard to tell apart, which our observability map should expose.
We deliberately do not propose truncating the 3×3 position Jacobian to "fix" vertical dilution of precision. That would only replace an unknown altitude with a guess. UTIL already carries a time-of-flight altitude laser that solves that problem more simply.
Research Program
The first step is diagnostic and needs ground truth. For each UTIL flight, we compute the residual table from motion capture. Within short windows, we then fit the anchor-level skew rank-2 term with SVD-RS style projection, using the degree normalization from the same paper because only some anchor pairs are measured at each instant. We report the fraction of residual energy captured by the anchor-level term. We also report how quickly the anchor-by-time table's singular values decay, which tells us how large $K$ needs to be. All of this is broken down by constellation and by centralized versus decentralized mode.
The second step is the estimator, which uses no ground truth. Over a sliding window of $T$ measurement epochs, we linearize around the current trajectory estimate. The unknowns are $3T$ position corrections plus $nK$ bias coefficients, solved in a Gauss-Newton loop. Each step solves $J\Delta = r$ (Jacobian $J$, step $\Delta$, residual $r$) by truncated pseudo-inversion. We keep singular direction $v_i$ only if its noise-amplified uncertainty is smaller than the plausible bias magnitude:
$$\sigma_i > \hat{s}/\beta,$$
where $\hat{s}$ is the robust residual noise scale estimated from the window and $\beta$ is a prior bias scale set from the hardware literature. This replaces a free tuning knob with two physically interpretable quantities. Measurements whose residual exceeds a decaying threshold are excluded before each solve, following TranSync. When the window slides, we update the truncated factorization with the Rayleigh–Ritz projection updates of Projection techniques to update the truncated SVD of evolving matrices, with periodic full refactorization to stop drift in the factorization itself. The window's bias estimate then feeds an otherwise standard error-state Kalman filter as a correction to the measurements.
The third step is analysis. For each window, the retained right singular vectors define the observable bias subspace. We relate its dimension to trajectory excitation, such as hovering versus wide sweeps and vertical motion versus planar motion. This turns "can we self-calibrate here?" into a quantity that can be measured.
Baselines
The naive reference is the UTIL benchmark pipeline as published: an error-state Kalman filter and batch estimation on uncorrected TDOA. UTIL reports an approximate positioning error of 10 cm in obstacle-free constellations #1–#3 and says performance "deteriorated significantly" in cluttered constellation #4. No reported result for constellation #4 was found on Lacuna, so we will rerun the released baselines ourselves and report both. Beating this reference means a lower 3D root-mean-square error and a lower 95th-percentile error on held-out flights, with per-axis errors reported separately.
The strongest competitor on the actual question is a bias-augmented Kalman filter. It is the same error-state filter, with one bias state per anchor and the gauge fixed by pinning a reference anchor. Bias process noise and initial covariance are tuned on the same training flights with the same tuning budget we give our two parameters $\hat{s}$ and $\beta$. This is the method a robotics practitioner would build, and it is effectively recursive ridge regression on the same unknowns. To beat it, our method must achieve lower error on held-out flights. Alternatively, it must achieve equal error with markedly lower sensitivity across the tuning grid, which is the advantage LoRanPAC reports for truncation over ridge in severely ill-conditioned least squares.
The upper reference is oracle calibration: per-anchor biases fitted from motion-capture residuals on training flights and subtracted on test flights. The fraction of the oracle's gain that our method recovers is the headline number. The ablation that isolates truncation replaces the truncated pseudo-inverse with Tikhonov regularization at a matched $\lambda = \hat{s}^2/\beta^2$, leaving all else identical. A second ablation removes the rank-2 anchor structure and estimates a free bias per anchor pair, testing whether the structure is what carries the gain.
Why not just…?
Why not just use ridge (Tikhonov) regularization instead of truncation? This is a real threat. For pure accuracy, ridge with a matched penalty may perform as well, and our claim does not rest on beating it. The matched-$\lambda$ ablation settles the accuracy question directly. The evidence that truncation might matter comes from a different domain, LoRanPAC's finding that ridge is more hyperparameter-sensitive. Our two reasons for preferring truncation are specific to this problem. Truncation gives an explicit, discrete observable subspace for the identifiability analysis. And it leaves unobservable bias directions exactly at the prior rather than partially fitting them to noise. If ridge wins on accuracy, we will report that and keep the spectral analysis as the diagnostic contribution.
Why not just add bias states to the Kalman filter? This is the bias-augmented baseline above. The filter will estimate weakly observable bias directions with whatever process noise it was tuned with. During hovering or planar motion, those states can absorb position error, especially vertical error, since that is where near-planar anchors make bias and altitude nearly interchangeable. Whether this happens on real data is an empirical question. We will answer it by comparing bias-state wander against the oracle biases, split by trajectory excitation.
Why not just cancel clocks in hardware, with relays or asynchronous message protocols? TDOA-MCM and the asynchronous scheme of Chiasson et al. remove clock offset and drift effectively; TDOA-MCM reports about 3 cm median error in simulation and under 10 cm in a classroom. But antenna delays and power-dependent delays survive these protocols, as Sidorenko et al. document. Our model targets exactly that residual. Running our estimator on top of drift-cancelled measurements shows whether any anchor-level bias remains to remove.
Why not just calibrate each station once? This is the oracle baseline. It requires motion capture or a calibration rig, and it does not track thermal or power-dependent changes during operation. If the oracle's static biases transfer perfectly across flights, the online method's main remaining value is deployability. The cross-flight transfer test measures exactly that.
Proposed Experiments
The decisive first check is the ground-truth diagnostic on UTIL. Before building anything, we pre-register a cutoff: if the anchor-level skew rank-2 term removes less than roughly a fifth of residual root-mean-square error in constellations #1–#3, the program stops. In that case the error is dominated by link-specific effects and the conclusion is negative but informative. A second diagnostic uses UTIL's controlled-obstacle identification recordings. It checks that NLOS delays behind cardboard, metal, wood, plastic, or foam show up in the sparse link term and not in the anchor term. If they leaked into the anchor term, the decomposition would be confounded.
The main experiment runs all baselines and our estimator on UTIL flight data across the four constellations and both TDOA modes, using a split by flight. We report 3D root-mean-square error, per-axis error with vertical error highlighted, 95th-percentile error, the fraction of the oracle's gain recovered, and sensitivity across the tuning grid.
A controlled simulation complements this, with bias, drift, and geometry known exactly. It follows the 10 m × 10 m, four-station setup of TDOA-MCM, with clock offsets up to 1 ns, where standard TDOA errors exceeded 50 cm. It adds ±20 ppm drift as in Chiasson et al., power-dependent per-anchor delays at the 0.3 m scale reported by Sidorenko et al., and both near-planar and well-spread anchor layouts. Here we can test the sharpest prediction. The dimension of the retained bias subspace should grow with trajectory excitation, and the error between estimated and true bias along the discarded directions should match the prior rather than noise. Finally, we sweep $\beta$ across an order of magnitude to quantify the resolution-versus-stability cost that the Fourier-extension work predicts for any truncation rule.
References
- UTIL: An ultra-wideband time-difference-of-arrival indoor localization dataset
- Ranking and synchronization from pairwise measurements via SVD
- Translation Synchronization via Truncated Least Squares
- Frames and numerical approximation
- Stable and accurate least squares radial basis function approximations on bounded domains
- Parameter selection and numerical approximation properties of Fourier extensions from fixed data
- Decawave UWB Clock Drift Correction and Power Self-Calibration
- Asynchronous Hyperbolic UWB Source-Localization and Self-Localization for Indoor Tracking and Navigation
- Projection techniques to update the truncated SVD of evolving matrices with applications
- LoRanPAC: Low-rank Random Features and Pre-trained Models for Bridging Theory and Practice in Continual Learning
- TDOA-MCM: A TDOA-based Multi-path Communication Mechanism for UWB Indoor Positioning