{"id":"17d6113f-00c2-40ff-91f3-40a12d682966","arxiv_id":"2507.15478","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A drone controller learns its own position-error distribution and uses it to re-weight a logical compliance map, yielding crash-free path choices in a real indoor testbed.","lead":"This paper presents a controller that combines logical rules for safe flight with a neural model of the drone's own tracking error, and shows that the combined system avoids obstacles that a rule-based baseline hits. It matters because it offers a way to make autonomous vehicles calibrate their confidence and plan more defensibly under uncertainty.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (10) as written marginalizes out the position variable that Eq. (11) optimizes over; doubt-calibrated compliance is formally undefined, so the reported steering advantage rests on an unwritten convolution.","rationale":"The paper's intended contribution is clear: augment a probabilistic-logic compliance field with a learned model of tracking error so that the planner keeps margin near obstacles. That idea is coherent, and the real-drone experiments are a meaningful existence proof. My stress-test focuses on the formal definition, because the entire downstream argument—Eq. (11), the landscapes in Figures 7–8, and the crash comparison—depends on Pθ being a function of the planned position x. Eq. (10) as printed is not such a function. This is an internal inconsistency, not a disagreement with consensus; it needs to be settled before the empirical comparison can be interpreted. The transferability concern raised by the reader is real and is partly acknowledged by the authors in Section V-E ('dedicated calibration for each new doubt feature'), but it is secondary: even a perfectly calibrated doubt model needs a well-defined way to enter the cost. If the authors confirm the convolution reading and release code showing that the implemented cost is the corrected Eq. (10), the conditional accept can stand; if the literal equation is what was optimized, the experimental result is unexplained by the paper's mathematics. No machine-checked proof or reproducible artifact is provided, so this formal check is the appropriate gate.","tokens_in":15351,"tokens_out":6642,"duration_ms":74126,"concrete_test":"Use the promised open-source CoCo code to instrument the exact computation of Eq. (10) for a fixed z: evaluate Pθ at two positions x1 and x2 that are separated by an obstacle. If the literal formula is used, both values are identical; if the implementation shifts the state by sampled errors, write the integrand explicitly as P(Ct|x+ε,z)δW(ε|θ) (or its Monte-Carlo estimate) and re-run the 45-flight comparison. Report whether CoCo's 0 crashes and ProMis's 11 crashes persist under the corrected formula, and confirm that the reported compliance landscapes in Figures 7–8 are generated by that corrected integrand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is Eq. (10): Pθ(Ct|x,z)=Ex∼δW(x|θ)[P(Ct|x,z)] = ∫P(Ct|x,z)δW(x|θ)dx. As written, x is both the argument of the left-hand function and the integration variable; after the integral the right-hand side no longer depends on the position part of x. Consequently, in Eq. (11) the term −α log Pθ(Ct|x,z) is position-independent for fixed z, and A* would minimize only β^T·J(x), so CoCo could not choose compliant paths at all. That contradicts the path switching reported in Section V-D and Figures 9–10. Section IV-D defines δW(x|θ) as 'a conditional probability distribution over the robot's state space x,' not as an error density centered at the commanded state, so this is not a harmless notation shortcut. The likely intended reading is a convolution with an error variable, e.g. Pθ(Ct|x,z)=∫P(Ct|x+ε,z)δW(ε|θ)dε, but that formula never appears and the training data in Figure 5 are L2 position errors, not absolute-state densities. The empirical 0-vs-11 crash result is suggestive, but it cannot validate an equation that, taken literally, removes spatial information from the cost; the implementation must be using an unwritten convolution, and the paper's central claim depends on that missing definition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Constitutional Controller (CoCo), a neuro-symbolic control architecture for unmanned aerial vehicles. CoCo combines a probabilistic first-order logic program (the Constitution), a Statistical Relational Map (StaR Map) of the environment, sensor/perception clauses, and a learned conditional normalizing flow called the 'doubt density' δW(x|θ). The doubt density is intended to represent the agent's tracking uncertainty under different controller tunings, velocities, and headings. The paper defines a doubt-calibrated compliance landscape Pθ(Ct|x,z) by integrating the compliance landscape with the doubt density, and then plans paths via A* on the cost −α log Pθ(Ct|x,z) + β^T·J(x). The experimental section reports a real-world Crazyflie study: in 45 flights, CoCo had zero crashes while the ProMis baseline crashed 11 times, all at high velocity, and CoCo additionally adapts its velocity when allowed. The paper argues that self-doubt calibration makes compliant agents safer than a baseline that uses the same constitution without the doubt mechanism.","tokens_in":15660,"tokens_out":6626,"duration_ms":76737,"significance":"The paper addresses a relevant problem—rule-compliant and safe autonomous UAV navigation under uncertainty—and the proposed combination of probabilistic logic programs with a learned, context-conditioned error model is genuinely interesting. If the central equations are corrected, the work would provide a concrete neuro-symbolic control mechanism with a real-world demonstration. The experimental effort is a clear strength: 45 physical flights with a Crazyflie, a plausible comparison to a state-of-the-art baseline, and a crash-free result for CoCo. The promise of open-source code is also a positive. However, the formal definition of the doubt-calibrated landscape is currently defective: the central equation, as written, removes the spatial dependence that the downstream planner relies on. Because the entire safety claim rests on this mechanism, the paper needs a substantive revision of the formulation and additional calibration evidence before it can be accepted.","major_comments":[{"comment":"Equation (10) does not define the object it claims to define. The left-hand side Pθ(Ct|x,z) retains x, but the right-hand side integrates over x with respect to δW(x|θ); for fixed z and θ, the integral is constant in x. Taken literally, the term −α log Pθ(Ct|x,z) in Eq. (11) cannot distinguish spatial positions, so the planner could not choose different paths based on compliance. This contradicts Figures 7–8, which show doubt-calibrated landscapes that clearly vary with x, and the path-switching behavior reported in Section V-D. The intended construction appears to be a convolution with an error variable, e.g., Pθ(Ct|x,z) = ∫ P(Ct|x+ε,z) δW(ε|θ) dε, but that formula never appears. Moreover, Section IV-D defines δW as a distribution over the robot's state space x, not over deviations from a commanded state. Please state the correct convolution, specify the domain of δW, and give the Monte Carlo estimator that was actually used in the experiments.","section":"Section IV-E, Eq. (10)"},{"comment":"There is a mismatch between the learning objective stated in Eq. (8) and the data described in Section V-C. Eq. (8) maximizes the log-likelihood of 'observed states' x^(i), but Section V-C says the CNF is fit to the L2 norm of the position error between desired and achieved positions, and Figure 5b plots scalar L2 error over time. A scalar L2 error is nonnegative and directionless; it cannot by itself determine the two-dimensional error density shown in Figure 6, and it cannot serve as the convolution kernel needed in Eq. (10). The paper should clarify whether δW(x|θ) models absolute state uncertainty or tracking-error uncertainty, what the actual training target is (2D error vectors or scalar L2 errors), and how the two-dimensional densities in Figure 6 were obtained. This is load-bearing because the doubt-calibrated landscape is only meaningful if the learned object has the right support and dimensionality.","section":"Section IV-D, Eq. (8) and Section V-C, Figure 5"},{"comment":"The doubt density is trained on 8-shaped flights at speeds shown in Figure 5 (0.6–1.0 m/s) and then applied to a different obstacle course at planned speeds that include 0.2 and 0.5 m/s, which are outside the training range. The paper provides no held-out calibration, coverage analysis, or evidence that the learned error model remains valid near the obstacles and under the actual path-following controller used in the mission. The central safety claim—that doubt-calibrated planning avoids crashes—depends on this transferability. Section V-E only notes that tuning may be context-dependent; it does not address calibration. Please add quantitative calibration checks (e.g., held-out error likelihoods, quantile coverage of the error density, or reliability diagrams) or explicitly scope the claim to conditions covered by the training distribution.","section":"Section V-C to V-D and Section V-E"}],"minor_comments":[{"comment":"There is a typo in the state description: 'veocity' should be 'velocity'.","section":"Section V-B"},{"comment":"The statement that 'choosing α=0 perfectly recovers the baseline behavior' is inaccurate: with α=0, Eq. (11) ignores the compliance landscape P(Ct|x,z), whereas the ProMis baseline is described as also minimizing a compliance term with weight α=2. Please rephrase the subsumption claim.","section":"Section V-E"},{"comment":"The notation J is overloaded: J(x) ∈ R^J with J ∈ N is confusing. Use a different symbol, e.g., K, for the dimensionality of the cost vector.","section":"Section IV-F, Eq. (11)"},{"comment":"The integral in Eq. (13) uses a filtering density p(xt|z1:t-1) but no online Bayesian update is specified in the paper; either define this quantity explicitly or move the equation to the discussion as a conceptual quantity.","section":"Section IV-G, Eq. (13)"},{"comment":"The captions say the doubt density is 'marginalized' for velocity or heading, but the notation in Eq. (10) does not show conditioning on θ; please align the notation so that the reader can see which doubt features are being marginalized in each figure.","section":"Figures 7 and 8"}],"recommendation":"major_revision","confidential_remarks":"The paper has a potentially interesting contribution and a real hardware demonstration, but the formal core needs to be repaired before the claims can be evaluated. The main risk is not the empirical crash counts but the undefined convolution in Eq. (10) and the mismatch between the stated training target and the actual data. If the authors can fix these points and add calibration evidence, the paper could become suitable for publication. No concerns about citation practices or novelty disclosure arose during review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is genuinely new: learning a conditional density of the agent's own control errors and folding it into a probabilistic-logic compliance landscape before planning. That is distinct from ProMis and the Constitutional Filter, and it is a sensible thing to try. The architecture is clean, and the Crazyflie experiment is a real asset; the 0-vs-11 crash comparison is striking, and the authors are honest about CoCo subsuming the baseline at alpha=0.\n\nThe soft spot is load-bearing. Equation (10) writes Pθ(Ct|x,z) = ∫ P(Ct|x,z) δW(x|θ) dx. On the right, x is integrated out, but the left still depends on x. Taken literally, the cost in Eq (11) cannot see position, so CoCo could not choose compliant paths at all. That contradicts Figures 8–10, so the implementation must be using a convolution with an error variable, e.g., P(Ct|x+ε,z) integrated against δW(ε|θ). That formula never appears, and Section IV-D defines δW as a density over the state space, not over errors, while the training data are L2 position errors. This is not a harmless notation shortcut; the paper's central claim rests on a missing definition.\n\nBeyond that, the transfer of the doubt model from 8-shaped training flights to the different obstacle course is unvalidated. There is no held-out calibration, no code or data yet, no confidence intervals, and only one baseline. The crash advantage is suggestive, but 45 flights is a small sample and the failure mode (high-velocity center path) is exactly what the cost landscape is designed to avoid.\n\nWho is this for? People working on neuro-symbolic control and safe UAV navigation. The idea is worth discussing and the empirical direction is right, but the formal derivation needs a rewrite before the result is usable. I would send this to peer review conditional on the authors fixing Eq (10), redefining the doubt density as an error distribution, and releasing code and data. As written, I would not cite the central claim, only the concept.","headline":"New and promising idea with a real drone experiment, but Equation (10) as written does not type-check and the central mechanism needs a corrected convolution.","tokens_in":16250,"tokens_out":1817,"would_cite":false,"duration_ms":20023,"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":"A drone controller that folds its own learned position-error distribution into its rule-compliance cost field avoided all collisions in 45 real flights, while a rule-only baseline crashed 11 times.","keywords":["constitutional controller","neuro-symbolic control","probabilistic logic programming","self-doubt","conditional normalizing flows","UAV compliance","uncertainty-aware path planning","statistical relational maps"],"falsifier":"Collect held-out position errors from actual CoCo missions on the obstacle course and compare them with the doubt density fitted on the figure-eight flights; if the density assigns low mass to the large errors that occur near yellow obstacles at 1.0 m/s, then Eq. (10) underestimates the true violation probability and the safety margin would not generalize. Directly retraining the doubt model on the mission course and rerunning the 45-flight comparison is the cleanest test, since the baseline-versus-CoCo gap should shrink if the transferred density was the cause.","tokens_in":15121,"feed_emoji":"🚁","tokens_out":9712,"duration_ms":93765,"temperature":0.7,"pith_summary":"The paper aims to show that a robot can be steered toward rules-compliance more safely if its planner explicitly accounts for how much the robot expects to miss its commanded position. It introduces the Constitutional Controller (CoCo), which combines a probabilistic-logic constitution of traffic rules with a learned conditional density of the agent's position error. The central move is to replace the raw compliance probability $P(C_t\\mid x,z)$ with its expectation under that error density, then plan paths trading off this doubt-calibrated compliance against travel cost. In 45 real drone flights through an obstacle course, CoCo suffered zero crashes while a rule-only baseline crashed on 11 flights, all at high speed. If true, this means uncertainty about one's own control accuracy can be converted directly into safer, still-compliant path choices without changing the rules.","feed_headline":"Doubt-aware drone controller completes 45 flights with zero crashes","feed_subtitle":"CoCo folds a learned error model into rule-compliance costs, steering around risky shortcuts that a rule-only baseline takes.","key_machinery":"The load-bearing object is the doubt-calibrated compliance landscape, the expectation of the logic-derived compliance probability under the drone's learned error density. A conditional normalizing flow built from a Masked Autoregressive Transform is fitted to position errors from practice flights, conditioned on controller tuning, speed, and heading; in the cost landscape this density is marginalized out by Monte Carlo, smearing the sharp rule boundaries into safety margins. The final path $\\tau^*$ minimizes the summed negative log doubt-calibrated compliance plus a weighted travel-cost term, and an A* planner on a discretized state grid (position and three speed levels) finds that path.","core_discovery":"The paper's central claim is that doubt-calibrated compliance, $P_\\theta(C_t\\mid x,z)=\\int P(C_t\\mid x,z)\\delta_W(x\\mid\\theta)\\,dx$ in Eq. (10), is what makes a neuro-symbolic controller safe. Here $P(C_t\\mid x,z)$ is the probability that the agent's current state and sensor reading satisfy the encoded constitution, and $\\delta_W(x\\mid\\theta)$ is a conditional normalizing flow describing how far the agent expects to be from its commanded state, given speed, heading, and controller tuning. Planning minimizes $\\sum_{x\\in\\tau}[-\\alpha \\log P_\\theta(C_t\\mid x,z)+\\beta^T J(x)]$ over candidate paths. The empirical result is that this integrated cost steers a small quadcopter away from a short, risky center path at high speeds: the rule-only baseline kept choosing that path and crashed 11 of 45 flights, while CoCo completed all 45 without a crash.","pith_inferences":["The same convolution trick does not depend on the logic-based constitution: any differentiable cost field could be smoothed by a learned error density, so the idea likely transfers to neural planners and occupancy-grid costs, though the paper does not test that.","The transferability of $\\delta_W$ is the fragile link: the paper fits doubt on figure-eight flights and applies it to a different course; retraining on the mission course and comparing crash counts would separate the value of doubt-calibration from the value of the specific training distribution.","Equation (10) is written as an expectation over $x$ while the left side still depends on $x$; making this well-defined requires specifying the marginalization properly, and a formal treatment of that integral would strengthen the claimed calibration semantics.","The observed crash pattern, with all baseline failures at high velocity, suggests the mechanism matters most near the speed where the drone's error distribution widens; adding doubt features such as battery state or wind, as the paper notes as future work, is a natural extension."],"forward_implications":["A rule-compliant planner gains a built-in safety margin: paths passing close to an obstacle at a speed where the drone's errors are large become less attractive than slower or more distant alternatives.","Because setting $\\alpha=0$ in Eq. (11) recovers the baseline behavior, CoCo is a strict generalization of the rule-only probabilistic mission planner rather than a separate pipeline.","The online compliance probability $P(C_t)$ gives a continuous risk signal during flight, which can trigger emergency routines or post-mission audits in addition to offline path choices.","When velocity is left free, CoCo uses its doubt model to slow down over constrained zones and speed up elsewhere, so safety and efficiency are balanced by the same mechanism."],"supporting_citations":[{"why":"Supplies the baseline probabilistic mission planner whose cost landscape CoCo doubt-calibrates and against which flight safety is compared.","marker":"[6]"},{"why":"Provides the Statistical Relational Map representation of noisy spatial relations (distance, over) used in the constitution.","marker":"[24]"},{"why":"Provides the Masked Autoregressive Transform density estimator behind the conditional normalizing flow for $\\delta_W(x\\mid\\theta)$.","marker":"[37]"},{"why":"Provides the hybrid relational probabilistic logic language in which the constitution clauses are expressed.","marker":"[17]"},{"why":"Provides the segmentation model used to detect obstacle geometries from drone imagery for the StaR Map.","marker":"[42]"}],"fun_headline_variants":["Doubt-calibrated drone steering logs 45 crash-free flights","Self-doubt helps drone obey safety rules, zero crashes in 45 flights","Neuro-symbolic controller uses doubt to keep drone from crashing","Doubt-aware drone beats rule-only baseline: 45 flights, zero crashes","CoCo: doubt-calibrated compliance keeps drone safe for 45 flights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The controller assumes that the position-error statistics learned on practice flights carry over unchanged to the actual obstacle course; if the drone's errors in the mission differ from what it learned, the claimed safety margin is not assured.","fun_headline_variants_meta":{"raw":{"variants":["Doubt-calibrated drone steering logs 45 crash-free flights","Self-doubt helps drone obey safety rules, zero crashes in 45 flights","Neuro-symbolic controller uses doubt to keep drone from crashing","Doubt-aware drone beats rule-only baseline: 45 flights, zero crashes","CoCo: doubt-calibrated compliance keeps drone safe for 45 flights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2830,"prompt_tokens":908,"completion_tokens":1922,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":1840}},"tokens_in":524,"tokens_out":1922,"duration_ms":18920,"temperature":1.0,"reasoning_tokens":1840,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:31:11.198812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect held-out position errors from actual CoCo missions on the obstacle course and compare them with the doubt density fitted on the figure-eight flights; if the density assigns low mass to the large errors that occur near yellow obstacles at 1.0 m/s, then Eq. (10) underestimates the true violation probability and the safety margin would not generalize. Directly retraining the doubt model on the mission course and rerunning the 45-flight comparison is the cleanest test, since the baseline-versus-CoCo gap should shrink if the transferred density was the cause.","supporting_citations":[{"cited_title":"Mission design for unmanned aerial vehicles using hybrid probabilistic logic programs,","cited_arxiv_id":null,"evidence_quote":"Supplies the baseline probabilistic mission planner whose cost landscape CoCo doubt-calibrates and against which flight safety is compared."},{"cited_title":"Star maps: Unveiling uncertainty in geospatial relations,","cited_arxiv_id":null,"evidence_quote":"Provides the Statistical Relational Map representation of noisy spatial relations (distance, over) used in the constitution."},{"cited_title":"Masked autoregressive flow for density estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the Masked Autoregressive Transform density estimator behind the conditional normalizing flow for $\\delta_W(x\\mid\\theta)$."},{"cited_title":"Probabilistic logic programming for hybrid relational domains,","cited_arxiv_id":null,"evidence_quote":"Provides the hybrid relational probabilistic logic language in which the constitution clauses are expressed."},{"cited_title":"SAM 2: Segment anything in images and videos,","cited_arxiv_id":null,"evidence_quote":"Provides the segmentation model used to detect obstacle geometries from drone imagery for the StaR Map."}],"review_version":1}