{"id":"8fe02b20-8bac-4c63-a52b-5f973ce52d11","arxiv_id":"2508.12166","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A one-step diffusion planner uses the spread of its own trajectory samples as a learned proxy for localization error, letting a robot choose which sensors to keep on and saving energy while still reaching goals.","lead":"This paper trains a fast path-planning model that also estimates how uncertain the robot is about its position, so the robot can switch sensors on and off to save energy. The authors tested it on a real unmanned boat and report lower sensing energy at similar goal-reach rates to always-on sensing.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trajectory-spread proxy for localization error is asserted calibrated under belief rasters and sensor masks that are out-of-distribution for a planner trained on accurate-state demonstrations; no calibration evidence is provided and the downstream SAC covariance bound inherits this risk.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing step: the calibration of the trajectory-spread proxy when conditioned on degraded belief rasters and sensor masks outside the training distribution. My stress-test agrees with that identification and sharpens the downstream consequence: because SAC uses the proxy directly in the reward for sensor selection, any miscalibration directly undermines the claimed covariance bound and the energy-savings result. The abstract provides no empirical calibration evidence, which is why I maintain a CONDITIONAL verdict rather than ACCEPT. I did not find a stronger or independent objection from the available material: the 10 ms runtime claim, the use of soft-actor-critic, and the marine deployment are all plausible and externally checkable if the calibration premise holds. The concrete test proposed here would settle the concern: it is a direct calibration check against ground-truth covariance under distribution shift, followed by a closed-loop verification that the proxy-based sensor selection actually enforces the stated covariance bound. If the full paper already contains such a study, the concern would be resolved; if not, the central claim lacks required support.","tokens_in":963,"tokens_out":2321,"duration_ms":30559,"concrete_test":"Hold out a test set of belief rasters with varying noise levels and sensor masks that deliberately span beyond the training distribution (e.g., masks with only 1–2 active sensors, heavily corrupted beliefs). For each instance, run B-COD to obtain the per-waypoint trajectory spread as a proxy variance, and compare it with ground-truth localization error measured against known simulator pose using a particle filter or analytic covariance rollout. Compute expected calibration error (ECE) from a reliability diagram binning proxy variance against empirical squared error, and also check a quantile-quantile plot. Then, in closed-loop simulation, feed the proxy to the trained SAC policy and verify that the actual pose-covariance bound is satisfied on these out-of-distribution masks. If ECE exceeds 0.1 or the bound is violated in more than 5% of episodes, the central claim fails; if calibration a","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that the spread of denoising trajectories provides a calibrated, differentiable proxy for expected localization error, and that this proxy is sufficient to guarantee a covariance bound during SAC-based sensor selection. The load-bearing condition is that the proxy remains calibrated for belief rasters and sensor masks that are rare or absent in the training distribution. The abstract states that the diffusion planner was trained on demonstrations that presuppose accurate state estimates; at deployment it is conditioned on degraded belief rasters and arbitrary sensor masks. If the proxy is calibrated only under near-optimal, high-quality beliefs, then the SAC reward signal is mis-specified: it will select minimal sensor subsets that appear safe according to an overconfident proxy but actually permit pose-covariance growth. The abstract gives no calibration procedure, no comparison against ground-truth covariance, and no analysis of distribution shift over masks/beliefs. This is not an internal inconsistency, but an unvalidated transfer assumption. If the full text contains a calibration study on held-out degraded beliefs and a closed-loop covariance-bound check, the concern would be resolved; based on the abstract alone, it remains the weakest point in the argument.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes B-COD, a one-step diffusion planner that is conditioned on a pose-belief raster and a sensor mask, and claims that the spread of its denoising trajectories provides a calibrated, differentiable proxy for expected localization error. On the basis of this proxy, a soft-actor-critic selects online the minimal sensor subset, with the stated goal of optimizing sensing energy while bounding pose-covariance growth. The authors report a 10 ms forward pass on real hardware and marine trials on an unmanned surface vehicle, claiming reduced sensing energy at matched goal-reach relative to an always-on baseline.","tokens_in":1216,"tokens_out":2477,"duration_ms":28384,"significance":"If the calibration claim holds, the paper would make a useful step toward closing the loop between data-driven trajectory generation and belief-space sensor scheduling, with a plausible real-time implementation. The reported marine trials are a strength, as is the explicit attempt to produce a differentiable, fast uncertainty signal without external covariance rollouts. However, the central contribution rests on an unvalidated transfer assumption: the trajectory-spread proxy is asserted to remain calibrated when conditioned on degraded belief rasters and sensor masks that may be out-of-distribution for a planner trained under accurate-state demonstrations. The significance is therefore conditional until calibration evidence is supplied.","major_comments":[{"comment":"The central claim, that 'the spread of its denoising trajectories yields a calibrated, differentiable proxy for the expected localisation error,' is load-bearing for the entire sensor-selection framework, yet no calibration procedure, no comparison against ground-truth pose covariance, and no quantitative uncertainty metrics are reported. Calibration is a strong quantitative property; simply noting that the spread is differentiable and monotone in some sense is not enough. Please provide calibration curves or correlation statistics of the proxy against true localization error, ideally over the same belief rasters and sensor masks used at deployment.","section":"Abstract"},{"comment":"The SAC objective is said to 'bound pose-covariance growth.' If that bound is evaluated with the same fitted diffusion proxy rather than with a ground-truth covariance estimate, the argument risks circularity: the model is used to define the quantity it is then claimed to bound. To resolve this, the manuscript should define the covariance bound explicitly and show a closed-loop check against actual pose covariance, e.g., from a recursive estimator, especially for sensor masks and belief conditions that were rare or absent during training.","section":"Abstract"},{"comment":"The abstract states that the diffusion planner presupposes accurate state estimates during training, yet deployment conditions it on degraded belief rasters and sensor masks. The proxy must transfer out of distribution for the sensor-selection result to be credible. There is no described analysis of how spread calibration varies with belief degradation or mask sparsity. A concrete test would be to measure calibration of the proxy on held-out bad beliefs and compare the sensor subsets chosen by SAC against those chosen under a ground-truth covariance oracle; without such evidence, the reported energy savings could be an artifact of overconfident uncertainty estimates.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract calls B-COD 'the first planner' of this kind. A sentence situating it against the closest prior work in data-driven belief-space planning would help substantiate that novelty claim.","section":"Abstract"},{"comment":"The marine trials are described only as reducing energy while matching goal-reach. Trial counts, error bars, and the definition of 'matching' should be reported in the full text so the reader can judge the significance of the hardware result.","section":"Abstract"},{"comment":"The phrase 'just-enough sensing' is not formalized in the abstract. A precise cost or constraint, such as a threshold on the predicted covariance, would clarify the optimization objective.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The reader's low confidence is justified. The manuscript as represented by the abstract makes a strong empirical claim ('calibrated') that is central to the method, but no calibration evidence is visible. My recommendation of major revision is based on the assumption that the full text may contain such a study; if it does not, the claim is unsupported and the contribution would need to be reframed. I would advise the editor to specifically request calibration results on out-of-distribution belief rasters and sensor masks, and a non-circular covariance-bound check, before considering this for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real integration claim, not a toy. The core idea—condition a one-step diffusion planner on a belief raster plus a sensor mask, then read off denoising trajectory spread as a differentiable localization-error proxy—is genuinely new relative to the prior work I know. And they back it with actual hardware: real-time USV trials, 10 ms planning, and reduced sensing energy at matched goal-reach. That is worth a serious look.\n\nWhat the paper does well: it sidesteps analytic covariance rollouts entirely, which is the brittle part of most belief-space planners. Putting the sensor mask into the conditioning and letting the planner's own variance stand in for expected localization error is a clever move, and the system-level claim is externally checkable in principle. The authors are also honest about the open problem they are attacking.\n\nWhere it gets soft: the abstract asserts that the trajectory spread is a calibrated proxy for localization error, but gives no calibration procedure, no comparison to ground-truth covariance, and no distribution-shift analysis over degraded belief rasters or sensor masks. That matters because the planner was apparently trained on demonstrations that presuppose accurate state estimates; at deployment you feed it uncertain beliefs and sparse masks. If the proxy is only calibrated near high-quality beliefs, the SAC reward is mis-specified and the covariance bound becomes a self-referential check: the bound is a function of the very proxy being validated. The stress-test note lands on this and, based on the abstract alone, I think it is the right concern. It is not an internal contradiction; it is an unvalidated transfer assumption that the full text needs to address with held-out calibration and a closed-loop covariance check.\n\nWho gets value: researchers working on active sensing, belief-space planning, or diffusion-based trajectory generation. The proxy idea alone, if properly validated, would get cited. But as submitted, the evidence in the abstract is not enough to convince me the central calibration holds.\n\nRecommendation: send it to peer review. The idea is novel enough, the hardware deployment is real, and the missing validation is exactly what a good referee should demand. This is a conditional accept-shaped paper, not a desk reject.","headline":"B-COD is a promising integration of one-step diffusion with sensor masking and a trajectory-spread uncertainty proxy, but the calibration claim is its load-bearing and unverified pivot.","tokens_in":1735,"tokens_out":1599,"would_cite":false,"duration_ms":20754,"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 belief-conditioned diffusion planner's trajectory spread serves as a calibrated localization-error proxy that drives online sensor selection.","keywords":["belief-space planning","one-step diffusion","sensor scheduling","localization uncertainty","trajectory planning","partially observable environments","unmanned surface vehicles","reinforcement learning"],"falsifier":"Give B-COD a held-out sensor mask and belief raster, compute the trajectory spread, and compare it with the posterior covariance from a particle filter on the same input; a divergence between the two for any selected mask would falsify the calibrated-proxy claim.","tokens_in":838,"feed_emoji":"🤖","tokens_out":10960,"duration_ms":101226,"temperature":0.7,"pith_summary":"The paper aims to establish that a one-step diffusion planner, conditioned on a raster of the robot's pose belief and a mask of which sensors are active, produces a calibrated, differentiable proxy for localization error simply through the spread of its denoised trajectory samples. If true, this proxy lets a soft-actor-critic policy choose the minimal sensor subset at each location online, optimizing sensing energy while holding pose-covariance growth within a task-specific bound. The resulting system, B-COD, is claimed to deliver a short-horizon trajectory, per-waypoint aleatoric variances, and the localization-error proxy in a single 10 ms forward pass, with no external covariance propagation. This would matter because it removes a major runtime bottleneck in belief-space planning and replaces brittle heuristic sensor-switching rules with a data-driven, uncertainty-aware policy, demonstrated on a real unmanned surface vehicle.","feed_headline":"One 10 ms forward pass picks the minimal sensors","feed_subtitle":"Trajectory spread acts as a localization-error proxy, letting a policy cut sensing energy on a real boat.","key_machinery":"The central machinery is B-COD, a one-step diffusion planner whose denoising spread is repurposed as a localization-error proxy. The model is conditioned on two inputs: a pose-belief raster, an image-like grid encoding the robot's estimated pose distribution, and a sensor mask, a binary vector specifying which sensors are currently active. The variance across the planner's output trajectories—the denoising spread—feeds a soft-actor-critic policy that selects the next sensor mask online. This design replaces external covariance propagation with a single differentiable forward pass, which is what makes the 10 ms runtime and real-time sensor scheduling possible.","core_discovery":"The paper's central claim is that the spread of a belief-conditioned diffusion planner's denoising trajectories is a usable, differentiable stand-in for the expected localization error. Concretely, B-COD conditions a one-step diffusion model on a pose-belief raster—an image-like encoding of the robot's current pose distribution—and a sensor mask indicating which sensors are active. The model returns a short-horizon trajectory, and the variance across denoised trajectory samples provides per-waypoint aleatoric uncertainty plus an overall proxy for localization error. The authors argue this proxy is calibrated and complete enough for a soft-actor-critic agent to learn which sensors to turn on","pith_inferences":["A natural test is to compare the proxy against ground-truth covariance from a particle filter across sensor-mask configurations the model did not train on; the paper gives no explicit calibration procedure, so this is the point most likely to limit deployment.","The same 'output spread as uncertainty' trick could apply to other generative planners and other downstream objectives, such as active perception or risk-bounded control, not just sensor selection.","The claim that the proxy is 'calibrated' may depend on the diversity of sensor masks and belief rasters seen in the demonstrations; extending the training distribution with sim-to-real variations would be a direct way to harden it."],"forward_implications":["External covariance rollouts disappear: one diffusion forward pass returns both the trajectory and an uncertainty estimate, so belief-space planning becomes a single differentiable step.","Sensor-switching heuristics are replaced by a learned policy that can directly optimize the trade-off between sensing energy and localization risk.","Per-waypoint aleatoric variances are available for free, enabling downstream risk-aware behaviors like cautious deceleration or replanning when uncertainty spikes.","Real-time operation on resource-constrained platforms becomes plausible for tasks with 10 ms planning horizons, as demonstrated on a marine vehicle."],"supporting_citations":[],"fun_headline_variants":["Diffusion trajectory spread picks the sensors you need","Just-enough sensing via one-step belief diffusion","10 ms diffusion plan, minimal sensors, same goal reach","Sensor mask from trajectory spread, no covariance rollouts","Belief-conditioned diffusion cuts energy on a real boat"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The whole sensor-selection scheme collapses if the spread of denoising trajectories is not a calibrated estimate of true pose uncertainty for belief rasters and sensor masks that differ from the training distribution.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion trajectory spread picks the sensors you need","Just-enough sensing via one-step belief diffusion","10 ms diffusion plan, minimal sensors, same goal reach","Sensor mask from trajectory spread, no covariance rollouts","Belief-conditioned diffusion cuts energy on a real boat"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000263,"raw_usage":{"total_tokens":1469,"prompt_tokens":810,"completion_tokens":659,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":583}},"tokens_in":554,"tokens_out":659,"duration_ms":6598,"temperature":1.0,"reasoning_tokens":583,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T19:35:30.427412+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give B-COD a held-out sensor mask and belief raster, compute the trajectory spread, and compare it with the posterior covariance from a particle filter on the same input; a divergence between the two for any selected mask would falsify the calibrated-proxy claim.","supporting_citations":[],"review_version":1}