Paying Map Contributors for What a Hidden Check Confirms

Collaborative maps, such as crowdsourced road maps, multi-robot maps of mines and tunnels, or 3D models stitched from many people's photos, are increasingly built from measurements that self-interested parties submit and get paid for. The tools that clean these measurements are geometric outlier rejection methods. They throw away matches between images or places that cannot all be explained by one rigid camera motion, and they assume the bad matches are accidents. A contributor who is paid per measurement has different incentives. It can copy geometry that is already in the shared map at no cost, duplicate its reports across fake identities, or invent plausible loop closures in places it never visited. A loop closure is a claim that "I am back where I, or another robot, was before." Copied data is perfectly consistent, so outlier rejection keeps it. Invented data in unmapped areas has nothing to contradict it, so outlier rejection keeps that too. The map's accuracy may look fine while the operator pays for nothing and the map quietly contains made-up regions.

We propose a payment rule that uses geometry as the referee. The operator keeps a small, secret, randomly placed set of trusted verification measurements. Each contributor is paid by how much its data improves the fit of the map to that hidden set, compared with the map built from everyone else's data. Copied data improves nothing and earns nothing. Duplicates earn nothing after the first copy. Fabricated geometry is punished whenever it lands where the hidden set happens to look, and contributors cannot know in advance where that is. Honest, effortful observation is the only reliable way to earn. The rule turns robust geometric estimation from a passive filter into the scoring device inside an incentive mechanism.

Core Claim

Let $V$ be the hidden verification measurements, let $\hat\theta_{-i}$ be the robust map estimate computed from all contributors except $i$, and let $\hat\theta$ be the estimate computed from everyone. Contributor $i$ receives a base fee $b$ plus

$$\pi_i = \lambda \left[ L(V;\hat\theta_{-i}) - L(V;\hat\theta) \right],$$

where $L$ is a truncated least-squares loss on the residuals of $V$ under a map estimate, and $\lambda$ is a payment scale. Truncated least squares means quadratic up to a cutoff and constant beyond it, the same robust cost used by graduated non-convexity. Two design choices carry the incentive argument. The first is that the reference estimate $\hat\theta_{-i}$ never contains contributor $i$'s own data. This is the "decouple the report from the price" principle of truthful mechanism design, carried over to geometry. The second is that $V$ is hidden and randomly located, so fabricated geometry cannot be tuned to agree with it.

In a locally linearized Gaussian model of the pose graph, which is the network of robot poses linked by odometry and loop-closure measurements, $\pi_i$ is a Clarke-style externality payment under a proper loss. We aim to prove that full-effort truthful reporting is a Bayes–Nash equilibrium there. Under the true nonlinear robust estimator, we expect only approximate truthfulness and will measure the deviation gain empirically. We state this plainly because the robust estimator is not a Bayes-optimal aggregator, so exact properness is not guaranteed.

Where This Sits

The geometric side supplies the estimator and the threat model. Graduated Non-Convexity for Robust Spatial Perception gives a deterministic robust solver that tolerates 70–80% random outliers. It produces per-measurement inlier weights, and we reuse it to compute $\hat\theta$ and $\hat\theta_{-i}$. DOOR-SLAM rejects false inter-robot loop closures by finding the largest set of mutually consistent ones, with consistency checked against each robot's odometry. Both methods assume errors are accidental. Neither has any notion that a consistent measurement might be worthless or invented.

Attacks and Defenses in Crowdsourced Mapping Services motivates the problem. It shows that scripted fake vehicles can outvote real drivers in Waze and divert them from a 0.4-mile route to a 0.7-mile one. Its defense is a physical-layer WiFi proximity check, not an incentive rule, and that check needs dense physical encounters between devices.

The mechanism side supplies the incentive logic, and it is closest to our work. Incentivizing Truthful Data Contributions in a Marketplace for Mean Estimation pays a primary contributor according to its distance from a single independent verifier sample. This is our spot-check structure, but only for a one-dimensional Gaussian mean. Truthful Linear Regression scores each agent's label against a model trained on a disjoint group. That is our leave-out decoupling, but it assumes the features are public and verifiable. In mapping, the coordinates themselves are what agents can fake. A Geometric Method to Construct Minimal Peer Prediction Mechanisms is geometric only in the space of beliefs. It does constrain us in one way: peer-prediction schemes without ground truth are known to reward agreement, and agreement is exactly what copying produces. This is why we anchor payments to a trusted hidden set instead of pure peer agreement.

Our proposal differs from all three in the same way. It handles reports that are multi-dimensional, constrained by camera and rigid-body geometry, and fully self-reported. It pays for marginal improvement against a hidden set, not for agreement with peers. Our budget concerns follow the frugality lens of Procurement Auctions with Predictions, which asks how much the buyer overpays relative to the cost of what it buys. That work gives no geometric results, so we use it only for the question of total spend.

Research Program

The theory step works in a linearized pose-graph model with Gaussian noise. Contributors there choose how much effort to spend, which sets the noise of their measurements, and what to report. They may report truthfully, copy the public map, fabricate in unvisited regions, or split their data across fake identities. We will derive conditions on the verification coverage probability $q$ and the scale $\lambda$ under which truthful full effort maximizes expected payment. We expect fabrication's expected payoff to scale like $-q\lambda$ times its expected residual penalty, and copying's to be exactly zero, because copied data does not change $\hat\theta$. We will also bound the total budget as a function of $q$, and characterize the operator's cost of verification.

The systems step implements the rule on top of graduated non-convexity in two settings. The first is multi-robot pose graphs built from DOOR-SLAM-style loop closures. The second is multi-contributor point-cloud registration. Computing the leave-one-contributor-out estimate $\hat\theta_{-i}$ for every contributor is the cost bottleneck. We will warm-start each $\hat\theta_{-i}$ from $\hat\theta$ and measure how much payment error this introduces.

The empirical step builds simulated strategic contributors, including best-response optimizers that search over attack families to maximize payment. We then measure how much any deviation gains over honesty.

Baselines

The naive reference point is pay-per-measurement with standard robust estimation. Contributors are paid a flat rate per submitted loop closure, and the map is cleaned by DOOR-SLAM's consistency-based rejection. This is what an operator would deploy today. It shows how much copying and fabrication pay when nothing discourages them, and how much of the map they corrupt.

The strongest existing competitor is pay-per-inlier. Contributors are paid only for measurements that survive the robust filter, using DOOR-SLAM's consistent-set selection or graduated non-convexity weights above 0.5. This is the natural "geometry already handles it" design. It is the closest existing method because it already ties money to geometric consistency.

No paper in the file reports results for strategic contributors, so there is no reported result on Lacuna for any baseline under strategic reporting. As an accuracy anchor in the honest-but-noisy regime, DOOR-SLAM reports Average Translation Error on KITTI00 dropping from 86.85 m with outliers to 8.00 m after rejection. Our mechanism must not degrade that honest-regime accuracy by more than a small margin.

The audit competitor is a binary spot-check. A contributor is paid a flat rate unless its data, on the hidden set, has a residual above a threshold, in which case it forfeits payment. This copies the structure of the mean-estimation verifier mechanism.

Three ablations isolate our key ingredients. The first makes $V$ public, which tests whether secrecy matters. The second computes the reference from all data including contributor $i$, which tests whether decoupling matters. The third pays by absolute fit to $V$ instead of marginal improvement, which tests whether the leave-out difference is what kills copying.

All methods share the same scenes, the same contributor population and attack strategies, the same verification budget (the same $q$ and number of trusted measurements), and the same total payment budget. Each method's scale parameter gets the same tuning budget on a held-out scene. We count our method as beating the competitors if, at equal budget and equal verification cost, three things hold together. Honest effort must be the best response, with a measured deviation gain near zero. Fabricated regions in the final map must be reduced severalfold relative to pay-per-inlier. Honest-regime trajectory error must stay within about 10% of the DOOR-SLAM rejection pipeline.

Why not just…?

Why not just rely on robust estimation and treat strategic reports as outliers? Robust estimators reject measurements that contradict the consensus. Copied data agrees with the consensus by construction, and fabrications placed in unmapped areas have no consensus to contradict. Neither is an outlier in the geometric sense. DOOR-SLAM and graduated non-convexity were evaluated only against accidental errors. The pay-per-measurement and pay-per-inlier baselines will show directly whether filtering alone leaves fabricated regions in the map and pays for copied data.

Why not just pay for inliers? Paying for survival of the filter rewards consistency, and consistency is the cheapest thing to fake. A lazy contributor copies; colluders fabricate a mutually consistent cluster of loop closures. This is the outvoting pattern Wang et al. documented in Waze. We expect pay-per-inlier to be the most exploitable baseline. If best-response attacks against it gain little, our proposal's motivation weakens, and we will report that.

Why not just run random binary audits? This is the real threat, because our rule is itself a form of audit. The difference is that a binary pass/fail audit pays copied data in full, since copies pass any consistency check, and gives no gradient toward effort. The marginal-improvement rule pays copies nothing and pays more for lower-noise data. If the binary audit matches our rule on deviation gain and map corruption at equal verification cost, the simpler mechanism should be preferred. The audit baseline is included specifically to settle this.

Why not just use physical attestation such as proximity handshakes or secure hardware? The WiFi defense reached an AUC above 0.98 in dense urban traffic, but it needs frequent physical encounters between devices. Sparse robot teams in GPS-denied tunnels rarely have such encounters. Attestation also proves a device was present, not that its geometry is honest. The two approaches are complementary, and attestation could reduce Sybil volume before our rule applies.

Proposed Experiments

We split the KITTI00 trajectory and the INTEL and CSAIL pose graphs into simulated contributors. A small trusted subset of loop closures and poses is held back as $V$. Contributors choose among truthful reporting, reduced effort (inflated noise), copying the current public map, fabricating loop closures consistent with odometry in unvisited segments, and splitting their data across fake identities. Multi-robot exploration scenarios in the ARGoS simulator add sparse-encounter conditions. Stanford Bunny registration with contributor-supplied correspondences tests the point-cloud case.

We measure four things. The first is Average Translation Error, or rotation and translation error for registration. The second is the fraction of map area supported only by fabricated data. The third is payment earned per unit of genuine information. The fourth is the empirical incentive gap: the payment gain of the best-found deviation over honesty, as a fraction of the honest payment.

The decisive checks are these. Copying must earn approximately zero under our rule. The incentive gap must fall toward zero as $q$ grows, at the rate our linearized analysis predicts. The public-$V$ and no-decoupling ablations must each reopen a profitable attack; if they do not, those ingredients are unnecessary. Our rule must dominate binary audits at matched verification cost, or we concede the simpler design. Finally, we trace the verification-cost versus incentive-gap curve, which tells an operator how much trusted data honesty costs.