Teams of robots, vehicles, or phones often locate themselves together. Each device measures radio distances to a few neighbors and to fixed beacons, and the group runs a shared filter so that position estimates spread through the network. Two things break this in practice. The first is sparsity: when a device can hear only one or two peers, its position is barely pinned down geometrically. The second is the gap between simulation and reality: real radio ranges contain reflections, blocked line-of-sight errors, and hardware biases that the Gaussian noise models used to design these filters never included. The two problems compound. A well-connected device can at least spot a bad range because it disagrees with other ranges. A device with one neighbor has nothing to cross-check against, so it cannot tell a corrupted range from a real move.
We propose that devices share compact, location-indexed statistics of how wrong their range predictions have been, which we call an interference map. They then use these statistics to set, online and with a calibration guarantee, how much each new range measurement should be trusted. A device that is currently isolated borrows the error statistics its peers recorded in the same area, so it can down-weight a bad range even when it has nothing local to compare against. The same calibrated range intervals, intersected with a motion prediction, give each device a position region that is meant to contain the truth at a chosen rate. That rate is what the experiments test hardest, and it matters most for the devices with the fewest neighbors.
Core Claim
In sparse cooperative positioning, the missing ingredient is calibrated knowledge of where the radio environment lies, more than a better estimator. That knowledge can be gathered without ground truth, because a device can always compare a new range measurement against the distance its filter predicted. That comparison is observable online. Online conformal inference turns a stream of such comparisons into an interval that covers future measurements at a target rate over the long run, with no distributional assumptions. We claim that pooling these comparison scores across nearby devices and across time, keyed by coarse location, restores good behavior in exactly the regime where local-only methods fail: devices with one or zero current neighbors inside zones where the line of sight is blocked. By good behavior we mean bounded error, few divergences, and position regions whose empirical coverage stays near target.
Relation to Prior Work
The closest single piece of work is the GNN-aided cooperative positioning method of Distributed Cooperative Positioning in Mobile Wireless Networks. It sends compact closed-form belief messages (a second-order Taylor approximation costing 34 arithmetic operations per iteration), refines them with a graph neural network, and adds an extended Kalman filter (EKF) so that temporal motion history keeps tracking stable when agents have only one or two neighbors. This paper motivates our choice of regime and supplies our estimator backbone. However, it assumes idealized random-walk mobility and clean ranging, and it reports no uncertainty guarantees. We keep its filter and change what goes into it: calibrated, location-dependent measurement trust. The complex-Laplacian method of Distributed localization in dynamic networks via complex laplacian attacks sparsity from the geometric side, cutting the neighbor requirement from three to two. It also ignores ranging corruption, so it addresses a different part of the problem from ours.
On the sim-to-real side, ULOC shows how severe real ultra-wideband (UWB) ranging errors are. UWB is a short-pulse radio used for precise distance measurement. On a 400 m by 200 m campus with 10 anchors, classical graph optimization degrades to errors "exceeding 10–20 meters", while a learned sequence model reaches about 2 m. ULOC is centralized, single-vehicle, specific to one environment, and gives no uncertainty. Its lesson that anchor silence and multipath carry information about location motivates our decision to index error statistics by place. MCD supplies the real UWB data and records a 70% segmentation drop across campuses. That result constrains us: it argues against a single offline-trained error model and for statistics learned online at each site.
Our uncertainty machinery reuses two methods. Conformal-AMC applies adaptive conformal inference to one wireless link and sets the adaptation rate from Doppler statistics. We take its online quantile update and its physically set step size and apply them to range residuals on a network. It works on a single link over synthetic 3GPP channels and notes that coverage conditioned on transient line-of-sight blockage is unsolved. Our degree- and zone-conditioned evaluation targets that open problem directly. Distributed Conformal Prediction via Message Passing shows that conformal quantiles can be agreed on across a graph by exchanging histograms of scores, and that sparse graphs such as rings and chains converge more slowly. We reuse its histogram exchange. We differ in three ways: scores come from a moving physical process, histograms are keyed by location, and they are carried forward in time. The last point addresses the slow convergence on sparse graphs, because a device inherits statistics collected before it became isolated instead of waiting for consensus it cannot reach.
The neighborhood-calibration idea of Authentication of Everything in the Internet of Things and the negative-observation vetoes of Dealing with Perception Errors in Multi-Robot System Coordination both rely on dense overlapping neighbors, and both papers note that they fail when neighbors are sparse. The interference map is our answer to that stated limitation, because it trades spatial overlap for temporal overlap. Finally, Consensus+Innovations Distributed Kalman Filter With Optimized Gains shows that a distributed filter's ability to track fast dynamics shrinks on weakly connected graphs. We treat this as a known ceiling that calibration cannot lift, and we measure how close to it our method operates.
Research Program
Each device $i$ runs the cooperative EKF of the GNN-aided method. On each link to a neighbor or anchor $j$, it computes a normalized residual $s_{ij} = |r_{ij} - \|\hat p_i - \hat p_j\|| / \hat\sigma_{ij}$. Here $r_{ij}$ is the measured range, $\hat p_i, \hat p_j$ are the filter's predicted positions, and $\hat\sigma_{ij}$ is the filter's predicted standard deviation of that range. No ground truth is needed. Each device keeps a small histogram of these scores for each coarse spatial cell it has visited, split by link type (peer or anchor). When two devices are in contact, they merge the histograms for cells near the sender's current estimate, using the histogram-consensus step of distributed conformal prediction, with a weight that decays with distance. Merged histograms persist, so a device entering a cell inherits whatever statistics earlier visitors left there.
From the local-plus-inherited histogram, the device forms an adaptive conformal quantile $q_{ij}$ for level $1-\alpha$. It updates the level online from its own coverage hits and misses, with a step size set from the device's speed and the cell's size, by analogy with the Doppler-derived rate of Conformal-AMC. The quantile does two jobs inside the filter. It rescales the measurement variance to $(q_{ij}\hat\sigma_{ij}/z_{1-\alpha/2})^2$, where $z$ is the standard normal quantile. It also rejects ranges whose residual exceeds $q_{ij}$ by a set factor. Each device's position region is the set of points that lie inside every calibrated range annulus and inside the motion-propagated region from the previous step. Computing it takes a few polygon intersections, which keeps the per-node cost within the budget that motivated the Taylor-message design.
The guarantee we can honestly claim is narrow. Adaptive conformal inference ensures long-run coverage of measured ranges, including under drift. Coverage of true positions is not guaranteed, because residuals mix estimation error with ranging error, and inherited histograms break exchangeability. Position coverage is therefore an empirical question, and the experiments are built to answer it separately by neighbor count and by line-of-sight condition.
Baselines
The do-nothing reference is the cooperative EKF with the fixed Gaussian range noise used in the original synthetic setup. It shows how badly a design tuned in simulation fails on realistic errors.
The strongest existing method is EKF-GNN-DCP from Distributed Cooperative Positioning in Mobile Wireless Networks. It is the only method in the file that targets distributed positioning at one to two neighbors per agent. It reports positioning error below 3 m for 80% of agents in dense settings, and it "maintained low error" in sparse settings where non-EKF methods spiked above 15 m. We will retrain its graph network and tune its noise settings on the same synthetic topologies and the same hyperparameter budget as our method. Its reported coverage is not a usable number, because there is no reported result found on Lacuna for coverage of positioning uncertainty regions by any method.
A tuned-inflation baseline replaces the fixed noise with the single scalar inflation that minimizes validation error on the trace-driven tier. This is the fair version of the heuristic-margin approach, whose link-adaptation analogue, 3GPP outer-loop link adaptation, reaches a 75.3% singleton rate against 82.4% for Conformal-AMC in Conformal-AMC. A robust-filter baseline swaps in a standard Huber-weighted, innovation-gated EKF with its threshold tuned on the same budget.
Two ablations isolate the key ingredient. The first uses local-only adaptive conformal inference with no histogram sharing. The second shares histograms globally with no location keys. If local-only matches the full method in the sparse bins, the interference map adds nothing. If global pooling matches it, location structure adds nothing.
A centralized EKF that sees every range at once serves as an upper reference for error, not as a competitor. For ULOC-style learned self-localization, the reported number is about 2.0 m RMSE on the campus route in ULOC. We cite it as context only, because it uses a different setup: a single tag with anchor ranges and no peer links.
Beating the strongest baseline means three things in the sparse, blocked-line-of-sight bin, under equal compute and tuning. The full method must cut the divergence rate (runs with error above 15 m) at least in half. It must match or improve RMSE. It must keep empirical position coverage within a few points of target where the robust filter and local-only calibration fall well short.
Why not just…?
Why not just inflate the measurement noise? A single inflation factor is either too loose in open areas, which wastes accuracy, or too tight in multipath zones, which lets outliers through. ULOC's report that graph optimization under standard noise assumptions exceeds 10–20 m on real campus UWB suggests the problem is spatial structure, not overall scale. The tuned-inflation baseline, broken down by line-of-sight zone, tests this directly. If it matches our method in the blocked-line-of-sight sparse bin, the idea fails.
Why not just use a robust, outlier-gated filter? This is the most serious threat, because Huber weighting and innovation gating are cheap and often work well. Their weakness is the one we target. A device with one neighbor has one residual, and a large residual is equally consistent with a corrupted range and with a real, unpredicted move, so a fixed threshold must guess. Borrowed statistics tell the device whether this cell usually produces large residuals. We expect the robust filter to be competitive when devices have many neighbors and to fall behind only at degree one or two. If it does not fall behind there, we will report that and drop the claim about accuracy, keeping only the coverage claim.
Why not just run conformal calibration locally at each device? Local adaptive conformal inference, the direct transplant of Conformal-AMC, needs many samples from each link and environment before its quantile settles; Conformal-AMC used 5,000 slots for initial calibration. A sparse device entering a new zone has a handful of samples. The local-only ablation measures how much coverage and error degrade in the first moments after a zone change. That window is where inheritance should matter, and it is where the comparison will be decided.
Why not just stream everything to a central server? When connectivity exists, a central filter is the better estimator, and we include it as a reference. The sparse regime is defined by connectivity that cannot be relied on. The file also documents that broadcasting all perceptual data costs two orders of magnitude more communication than selective exchange Dealing with Perception Errors in Multi-Robot System Coordination. We report bytes per device per step so that the cost of the interference map is visible next to this reference.
Proposed Experiments
The evaluation has three tiers of realism. The first tier uses the synthetic Monte Carlo topologies of the GNN-aided positioning work: 300 m to 900 m grids with 30 to 60 agents, and density swept down to mean degree below two. Ranges there have Gaussian noise, so this tier checks that calibration costs nothing when the design assumptions hold.
The second tier keeps those topologies and replaces the range noise with real UWB range errors. We obtain the errors by subtracting ground-truth distances from the B-spline-deskewed MCD trajectories from the measured UWB ranges. We then assign them to synthetic zones by matching line-of-sight state and distance, and we hold out whole MCD campuses (NTU, KTH, TUHH) between tuning and test. This tier is the central sim-to-real test: every method is tuned on one campus's error character and tested on another's.
The third tier uses Wi-GATr as a differentiable channel surrogate over WiPTR indoor scenes to generate range errors specific to each site. Its reported 70% error reduction against calibrated ray tracers on real DICHASUS measurements makes it the most credible available generator of structured multipath. We run it offline only, and it is never part of the per-node computation.
Metrics are position RMSE, divergence rate, empirical coverage of position regions and region area, empirical coverage of range intervals, and bytes exchanged per step. Every metric is reported separately by current neighbor count (zero, one, two, three or more) and by line-of-sight condition.
The decisive check is the degree-one, blocked-line-of-sight bin on held-out campuses in the second tier, compared against the robust filter and the local-only ablation. A secondary check measures time-to-recovery of coverage after a device enters a zone for the first time, compared with entering a zone a peer has already mapped. If this gap is absent, the core claim that inherited statistics carry information through sparsity is false.
We will also test directly the tracking-capacity limit from the consensus-plus-innovations analysis. We raise agent speed until every method diverges and check that our method's breaking point moves closer to the centralized reference's.
References
- Distributed Cooperative Positioning in Mobile Wireless Networks: A GNN-Aided Joint Model- and Data-Driven Framework With High-Accuracy Closed-Form Message Representation
- Distributed Localization in Dynamic Networks via Complex Laplacian
- ULOC: Learning to Localize in Complex Large-Scale Environments with Ultra-Wideband Ranges
- MCD: Diverse Large-Scale Multi-Campus Dataset for Robot Perception
- Distribution-Free Adaptive Modulation and Coding via Online Conformal Inference under Non-Stationary Fading
- Distributed Conformal Prediction via Message Passing
- Authentication of Wireless Devices in the Internet of Things: Learning and Environmental Effects
- Dealing with Perception Errors in Multi-Robot System Coordination
- Consensus+Innovations Distributed Kalman Filter with Optimized Gains
- Differentiable and Learnable Wireless Simulation with Geometric Transformers