{"id":"f28681d0-4d95-4789-a978-120e7ffa7e71","arxiv_id":"2412.18356","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"StaR Maps turn uncertain map data into probability distributions for spatial queries, allowing planners to ask how likely a statement about the environment is.","lead":"This paper introduces StaR Maps, a way to store map data together with statistical information about how accurate each feature is, and to answer spatial queries with probabilities instead of fixed answers. A smart generalist might read it because autonomous vehicles and drones need to reason about uncertain maps, and this work shows one way to combine crowd-sourced map data with probabilistic reasoning.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"StaR Maps store each spatial relation as an independent marginal, so the demonstrated first-order logic queries over conjunctions have no joint distribution and can return miscalibrated probabilities.","rationale":"The reader's weakest assumption—availability of per-feature error parameters—is real and limits the validation, but the paper explicitly acknowledges it as a simplification and points to [42] for online estimation on road segments. The joint-distribution omission is different: it is not acknowledged, it directly affects the paper's central reasoning claim, and it survives even if perfect error parameters are supplied. The experiments in Sec. IV only evaluate interpolation error (MAE) and visual fields; they never check whether any logical query returns the probability actually induced by the uncertain map. The proposed test is cheap, decisive, and uses only the released code and the paper's own setup. Because the demonstrated logical query mechanism can produce probabilities that are not those of the underlying uncertain map, the central claim is not supported as written. A revised version with explicit joint estimation, or with a clearly stated and empirically validated independence assumption, could change this assessment.","tokens_in":11394,"tokens_out":4309,"duration_ms":44865,"concrete_test":"Use the released ProMis code on a real OSM region with the paper's N(0, diag(10m, 10m)) error model. For a grid of query points x, compute from the N sampled maps the empirical joint probability P(distance(x, road) < 15 AND distance(x, pilot) < 250) and compare it with the product of the StaR Map marginals, or with the probability returned by the Listing 1 program. Report the maximum absolute and relative deviation across the grid. If the deviation is negligible, the independence assumption may be benign in the tested regime; if it is large, the reasoning pipeline is miscalibrated and needs a joint model (e.g., multivariate moment matching or retained samples).","verdict_should_be":"REJECT","load_bearing_attack":"The central claim is that StaR Maps provide probabilistic answers to spatial queries. In Sec. III.C, each spatial relation is fit independently by moment matching over the sampled maps W, and the scalar fields ui,j and vk,l are stored separately. Listing 1 then combines these marginals in a hybrid probabilistic logic program, e.g., airspace(X) :- distance(X, road) < 15, distance(X, pilot) < 250. This conjunction requires P(distance(X, road) < 15 AND distance(X, pilot) < 250), but the StaR Map retains only the two marginals, with no covariance or shared latent variable. The sampled maps W do contain the joint information, but it is discarded after moment matching. Unless independence is assumed—which is not stated and is false for two distances measured from the same randomly translated map—the probability returned for the conjunction is not the probability induced by the uncertainty model. This is not an experimental gap: the demonstrated reasoning mechanism is unsound for composite queries, so the central claim is unsupported for the paper's own flagship example. This concern is orthogonal to the reader's error-parameter concern; even with perfect per-feature error parameters, the joint mismatch remains.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Statistical Relational Maps (StaR Maps), a representation for uncertain geospatial data. Maps are augmented with per-vertex error parameters to form Uncertainty Annotated Maps, from which N random map instances are sampled. For each query point and feature type, categorical relations (e.g., over) and quantitative relations (e.g., distance) are summarized via moment matching into Bernoulli and Gaussian distributions, and scalar fields over the navigation space are approximated by linear interpolation or Gaussian Process regression. The authors demonstrate the pipeline on OpenStreetMap data with a synthetic isotropic translation error and show that GP-based construction reduces interpolation error more efficiently than grid sampling. They also show a hybrid probabilistic logic program that queries spatial constraints over the resulting scalar fields.","tokens_in":11602,"tokens_out":6889,"duration_ms":59683,"significance":"If validated, the representation would be a useful bridge between raw GIS data and statistical relational reasoning: it gives a uniform interface for probabilistic spatial predicates and proposes a scalable sampling-plus-regression construction. Strengths include clear formal definitions, an open-source implementation, and honest acknowledgment of limitations (e.g., the synthetic error model and the need for better distribution families). The main significance is limited by the fact that the evaluation is self-consistency on a single synthetic error model; the paper does not yet demonstrate calibrated probabilities on real uncertainty, nor does it validate composite query probabilities.","major_comments":[{"comment":"The central claim that StaR Maps provide probabilistic answers to first-order logic spatial queries is unsupported for composite queries. Each spatial relation is stored as an independent marginal scalar field (ui,j, vk,l), and the sampled maps W are discarded after moment matching. Listing 1 includes the conjunction airspace(X) :- distance(X, road) < 15, distance(X, pilot) < 250; its probability is not determined by the two marginals unless independence is assumed, and the paper neither states nor justifies that assumption. Since the two distances are measured on the same randomly translated map, they are typically dependent, so the probability returned for this conjunction is generally not the probability induced by the Uncertainty Annotated Map. The demonstrated reasoning mechanism is therefore unsound for the paper's flagship example. Please either preserve the joint information (e.g., evaluate composite queries by Monte Carlo on W, or store cross-covariances) or explicitly restrict the claims to single-relation queries with a stated independence assumption.","section":"Section III.C, Definition 1.3, Listing 1"},{"comment":"The experimental evaluation supports only internal self-consistency under a single synthetic error model, not calibrated probabilities on real map uncertainty. All OSM features are given the same Gaussian translation error N(0, diag(10 m, 10 m)), which the authors call a 'stark simplification.' The conclusion acknowledges that general error-parameter estimation and non-Gaussian relation models are future work. Because the correctness of the resulting probabilities is conditional on the input error parameters and the assumed distribution families, the experiments do not establish the abstract's claim of an 'honest view of the data's accuracy.' There is also no baseline comparison (e.g., a deterministic map query) and no error bars on the MAE curves in Figures 7 and 9. Please add experiments with heterogeneous, correlated, or non-Gaussian error models and calibration checks, or temper the claims to the synthetic setting.","section":"Section IV.A and Conclusion"},{"comment":"The moment-matching step assumes a Gaussian distribution for distance relations and a Bernoulli distribution for over, but this parametric assumption is not validated. The distance to the closest road under Gaussian map translation is generally not Gaussian, yet probability queries such as P(D > 30 m) in Figure 4(c) are computed from the Gaussian CDF. Figure 5 overlaid a Gaussian on a histogram but reports no goodness-of-fit or calibration statistic. The conclusion mentions that Gaussian models are 'not general tools,' but this is load-bearing for the central claim: a misspecified family leads to miscalibrated probabilities. Please provide a quantitative assessment of the fit, or show that the query probabilities are robust to the choice of family.","section":"Section III.C and Figure 5"}],"minor_comments":[{"comment":"The figure caption ends with a stray '43' that appears to be a formatting artifact and should be removed.","section":"Figure 7"},{"comment":"The equation for the translation sample uses the symbol ⃗βi, but the annotator function is defined as b(⃗ v) = ⃗β; the subscript is otherwise unused and should be removed for consistency.","section":"Section III.A"},{"comment":"The term 'Star Map' appears once and should be written 'StaR Map' for consistency with the rest of the paper.","section":"Section IV.B"},{"comment":"The syntax 'distance(x, building) ˜ normal (20, 0.5)' contains a spacing and tilde typo, and the feature type 'pilot' is used without explaining how it is represented in the map data.","section":"Listing 1"},{"comment":"The integer e in the codomains [0,1]^e and R^e is not defined; it should be stated as the number of parameters of the relation distribution.","section":"Definition 1.3"},{"comment":"Figure 3 illustrates translation, scaling, and rotation error models, but the experiments only exercise translation; the authors should state whether the implementation supports the other transformations.","section":"Figure 3"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is likely appropriate for a specialized venue on probabilistic robotics or spatial reasoning, but the current validation is too thin for a journal-level claim of probabilistic geospatial reasoning. The joint-distribution issue in Section III.C is the most serious and should be addressed before resubmission; the experimental section also needs to move beyond self-consistency on one synthetic error model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The honest take: StaR Maps is a well-scoped integration of existing ingredients—uncertainty-annotated maps, Monte Carlo sampling, moment matching, and GP interpolation—into a field-based representation that can answer single-relation queries. The package is new, the writing is clear, and the authors ship code and are upfront about their simplifications. For a single spatial relation (distance to nearest road, say), the pipeline is coherent and the interpolation experiments are sound. The GP confidence-guided refinement is a sensible engineering improvement over uniform raster sampling.\n\nThe problem is in the reasoning section. Each spatial relation is fit independently and stored as a marginal scalar field. Listing 1 then combines these marginals in a hybrid probabilistic logic program, e.g., airspace(X) :- distance(X, road) < 15, distance(X, pilot) < 250. Evaluating this conjunction requires the joint probability of both conditions. The StaR Map discards that joint information after moment matching; the sampled maps W contain it, but only the marginals survive. Unless independence is assumed—which is false here, since two distances measured from the same randomly translated map are correlated—the returned probability is not the probability induced by the uncertainty model. This is not an experimental gap; it is a theoretical mismatch between the representation and the advertised reasoning capability. The paper's own flagship example is unsupported as stated.\n\nOther soft spots are less severe. The experiments annotate every OSM feature with the same synthetic Gaussian translation error, which the authors admit is a stark simplification. No method is given for estimating error parameters on buildings, parks, or other feature types, so the utility claim rests on an assumption that is only gestured at for road segments. There is also no comparison to ProMis [26] or any prior probabilistic map representation; the MAE curves validate interpolation against a dense raster generated from the same error model, which checks self-consistency but not external validity.\n\nWho should read this: people working on uncertain maps for navigation will find the representation and the GP-based field construction useful, and the open-source code is a concrete artifact. As a probabilistic reasoning framework, it is not ready as stated. I would send it to peer review because the core idea is worth engaging with, but the authors need to either store the joint structure (e.g., shared latent map sample, covariance fields) or restrict claims to single-relation queries. With that change, the paper would be substantially stronger.","headline":"StaR Maps is a clean integration of known pieces into a queryable uncertainty-annotated map, but the flagship probabilistic reasoning over conjunctions is not supported by the stored marginals.","tokens_in":12111,"tokens_out":1640,"would_cite":true,"duration_ms":17932,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"StaR Maps turn per-feature map uncertainty into probability distributions over spatial relations, returning probabilities instead of single geometric answers.","keywords":["statistical relational maps","uncertainty annotated maps","spatial relations","probabilistic logic programming","map uncertainty","Gaussian process interpolation","autonomous navigation","crowd-sourced maps"],"falsifier":"Compare a StaR Map's predicted probabilities with empirical frequencies on a region where ground-truth positions are known independently: for many query points, record the predicted probability that the distance to the nearest road exceeds 30 meters and compare it with the frequency observed after applying the same error model to surveyed geometry. Systematic divergence where per-feature accuracy varies would show that the single synthetic error model miscalibrates the map's outputs.","tokens_in":11184,"feed_emoji":"🗺️","tokens_out":11453,"duration_ms":97497,"temperature":0.7,"pith_summary":"This paper introduces Statistical Relational Maps (StaR Maps), an environment representation that treats every map feature as carrying a positional error model and turns that uncertainty into probability distributions over spatial relations. The authors argue that common digital maps, especially crowd-sourced ones, omit accuracy information, and that navigation and planning systems should be able to ask for the probability that a spatial fact holds, such as a point lying over a park or farther than 30 meters from a road. A StaR Map is built by sampling many perturbed versions of an uncertainty-annotated map, fitting categorical and continuous distributions to the resulting relation statistics, and interpolating the parameters into scalar fields over navigation space. Experiments on real-world crowd-sourced map data show that the construction scales with Gaussian-process interpolation and that the resulting fields support probabilistic first-order logic queries.","feed_headline":"StaR Maps turn map errors into probability fields for spatial queries","feed_subtitle":"Every distance and location query returns a probability based on the map's own accuracy, not a single guess.","key_machinery":"The machinery has three linked pieces. First, an Uncertainty Annotated Map (UAM) augments each vertex of a map with translation parameters and transformation parameters, so a stochastic error model can generate many alternative map instances via affine perturbations such as translation, rotation, and scaling. Second, spatial relations between a point, a feature type, and a sampled map are evaluated across all instances and summarized by moment matching, meaning the distribution's parameters are set from the sample statistics: a Bernoulli for categorical relations such as over (is the point inside or above a feature of the given type) and a Gaussian for quantitative relations such as distance to the closest feature of the given type. Third, the relation parameters are turned into functions on continuous space by interpolation, and a Gaussian Process variant selects new sampling locations by its own predictive variance, making the field construction scalable and providing a natural stopping criterion. The sampling from the UAM is what connects raw map-error assumptions to queryable probability statements; without it, the scalar fields would reduce to deterministic distance or occupancy maps.","core_discovery":"On the paper's own terms, the central claim is that a StaR Map provides a unified representation that takes map-related error parameters, processes them through stochastic sampling, and stores the processed statistics so applications can reason probabilistically about spatial facts. Concretely, for each point in navigation space, the categorical relation over(feature type) is modeled by moment-matching a Bernoulli distribution and the quantitative relation distance(feature type) by a Gaussian, using N sampled maps generated from an Uncertainty Annotated Map, i.e., a map whose vertices carry translation and transformation statistics. These relation parameters are approximated as continuous scalar fields, with a Gaussian Process variant that samples additional points where its confidence is low. The authors demonstrate that the resulting fields answer logical queries, such as a simple airspace rule defined over park and road relations, with probabilities rather than crisp geometric verdicts.","pith_inferences":["The paper leaves implicit a natural next step: estimate the per-feature error models online by comparing crowd-sourced geometry against higher-accuracy surveys, extending the road-segment method to buildings, parks, and other feature types.","If real relation statistics are skewed or heavy-tailed, moment matching to Gaussian and Bernoulli families will miscalibrate; testing mixture or beta models would show how much calibration is lost by the fixed distributional choice.","A promising untested connection is feeding detector covariances from neural perception into the Uncertainty Annotated Map, so sensing uncertainty and map uncertainty are answered by the same query language.","The GP's confidence field could double as a safety monitor: a planner could refuse to act in regions where the variance of the relation statistics exceeds a threshold, since those are exactly the places the map cannot support a reliable decision."],"forward_implications":["If StaR Maps work as claimed, any point query against a map can return a probability rather than a single geometric answer, such as the chance that the distance to the nearest road exceeds 30 meters.","The Gaussian-process construction makes it practical to build these probability fields over wide urban areas, adding sample points where uncertainty is high instead of covering the whole space with a uniform grid.","Hybrid probabilistic logic programs can consume StaR Map fields directly, so regulatory or mission constraints phrased as logical rules become queries with probabilistic answers.","The representation extends the earlier mission-design approach by putting a probabilistic regression model on top of the sampling process, so the map itself reports where its relation statistics are least certain.","Because the uncertainty annotation is per vertex, the same pipeline applies to roads, buildings, parks, and other tagged features once suitable error parameters are supplied."],"supporting_citations":[{"why":"Supplies the online error-parameter estimation method for road segments that the paper relies on when uncertainty annotations are not provided by the map.","marker":"[42]"},{"why":"Provides the hybrid probabilistic logic query interface over navigation spaces that StaR Maps extend with probabilistic regression and confidence-guided refinement.","marker":"[26]"},{"why":"Supplies the real-world crowd-sourced map data and the observation that such maps omit positional accuracy information, motivating the uncertainty annotation.","marker":"[34]"},{"why":"Defines the distributional-clause hybrid probabilistic logic programming language used to express and query the knowledge contained in a StaR Map.","marker":"[27]"}],"fun_headline_variants":["StaR Maps turn map errors into spatial probability fields","Probabilistic maps reveal uncertainty in every road query","Spatial relations with honest probabilities from map data","Geospatial queries now return probabilities, not guesses"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on having trustworthy estimates of how far each map feature might be from its true position; in the experiments every feature is given the same synthetic Gaussian error, so if real errors differ by feature, are correlated, or are non-Gaussian, the output probabilities are not dependable.","fun_headline_variants_meta":{"raw":{"variants":["StaR Maps turn map errors into spatial probability fields","Probabilistic maps reveal uncertainty in every road query","Spatial relations with honest probabilities from map data","Geospatial queries now return probabilities, not guesses"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000168,"raw_usage":{"total_tokens":1235,"prompt_tokens":892,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":508,"completion_tokens_details":{"reasoning_tokens":280}},"tokens_in":508,"tokens_out":343,"duration_ms":3641,"temperature":1.0,"reasoning_tokens":280,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:45:57.803843+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare a StaR Map's predicted probabilities with empirical frequencies on a region where ground-truth positions are known independently: for many query points, record the predicted probability that the distance to the nearest road exceeds 30 meters and compare it with the frequency observed after applying the same error model to surveyed geometry. Systematic divergence where per-feature accuracy varies would show that the single synthetic error model miscalibrates the map's outputs.","supporting_citations":[{"cited_title":"Error decomposition for hybrid localization systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the online error-parameter estimation method for road segments that the paper relies on when uncertainty annotations are not provided by the map."},{"cited_title":"Mission Design for Unmanned Aerial Vehicles using Hybrid Probabilistic Logic Programs,","cited_arxiv_id":null,"evidence_quote":"Provides the hybrid probabilistic logic query interface over navigation spaces that StaR Maps extend with probabilistic regression and confidence-guided refinement."},{"cited_title":"Openstreetmap: User-generated street maps,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world crowd-sourced map data and the observation that such maps omit positional accuracy information, motivating the uncertainty annotation."},{"cited_title":"Probabilistic logic program- ming for hybrid relational domains,","cited_arxiv_id":null,"evidence_quote":"Defines the distributional-clause hybrid probabilistic logic programming language used to express and query the knowledge contained in a StaR Map."}],"review_version":1}