{"id":"f6800883-8255-4070-bc96-1e8d1f6ff020","arxiv_id":"2412.10706","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":9,"one_line_summary":"SHIFT combines semantic-aware speed control with incremental local replanning to make robot coverage more uniform in dynamic 3D environments.","lead":"This paper presents SHIFT, a coverage planning system for cleaning and agricultural robots that combines semantic maps, terrain-following paths, and speed modulation based on Gaussian fields. The authors report better coverage uniformity and lower planning time than several existing planners in simulations and on a robotic vacuum.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (17) evaluates the 2D Gaussian disk integral as a 1D CDF, so the RFICP speed law in Eqs. (18)-(19) does not follow from the stated diffusion model; this invalidates the derivation behind the coverage-uniformity claim.","rationale":"The paper's central claim is state-of-the-art coverage uniformity and adaptability, and the proposed mechanism is the RFICP speed law: a Gaussian diffusion model in Eq. (16) is inverted into dwell time and speed in Eqs. (18)-(19). The reader's weakest assumption identifies exactly the step where this inversion goes wrong: the disk integral of the Gaussian kernel is evaluated with a 1D Gaussian CDF. I recomputed the integral and find that the stated equality in Eq. (17) is false for the kernel as written, and also false under the natural 2D-normalized correction. Since the speed law is the only formal support for the coverage-uniformity advantage, this is a load-bearing error. The secondary issue of pulling t(p) out of the p' integral reinforces the conclusion that the derivation does not establish the speed law, even if the normalization were repaired. The paper includes real experiments and a system, but no machine-checked proof or complete reproducible code is supplied, and the experimental tables lack error bars; these factors do not rescue the derivation. Because the central speed-allocation law is unsupported as written, I see no reason to change the reader's REJECT verdict. A corrected derivation and re-run experiments could lead to a different outcome, but that is not the current submission.","tokens_in":10286,"tokens_out":5472,"duration_ms":53440,"concrete_test":"Recompute the integral in Eq. (17) symbolically or numerically for representative (σ,R), for example σ=1 and R=3, comparing the exact 2D disk integral of Eq. (15) with 2f(R/σ)-1. Then re-derive t(p) and v(p) using the correct expression and compare with Eqs. (18)-(19); if the functional form changes materially, the RFICP speed law evaluated in the experiments is not the one derived from the stated model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing mathematical step is the inversion of the Gaussian diffusion model in Section V. Equation (17) asserts that the 2D disk integral of the kernel in Eq. (15) equals 2f(R/σ)-1, where f is the 1D standard Gaussian CDF. With the kernel written as a 1D-normalized Gaussian in two spatial dimensions, the actual integral over the disk of radius R is sqrt(2π) σ [1 - exp(-R^2/(2σ^2))]; if the kernel were intended to be 2D-normalized, the integral would be 1 - exp(-R^2/(2σ^2)). Neither equals 2f(R/σ)-1. Consequently Eq. (18) for t(p) and Eq. (19) for v(p) do not solve Eq. (16), and the speed-allocation law that is supposed to produce uniform semantic coverage is unsupported. There is also a conflation in the same step: t(p) is pulled out of the p' integral as if the dwell time at the robot position p' were a function of the evaluation point p; even a corrected disk integral would not by itself justify the closed-form inversion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The SHIFT planner paper integrates semantic mapping, terrain-adaptive coverage path generation, and local obstacle avoidance into a single navigation framework. It proposes RFICP, which models the spatial influence of a cleaning action as a Gaussian diffusion field and derives a speed allocation law from an inverted coverage integral, and IKD-SWOpt, which uses an incremental KD-tree distance field with A* initialization and sliding-window optimization to avoid dynamic obstacles. Experiments on a robot vacuum cleaner and in simulation report improved coverage completeness, uniformity, energy efficiency, and planning time relative to baselines. The central theoretical step is the closed-form inversion of a Gaussian disk integral in Eq. (17), and that step is mathematically incorrect as written.","tokens_in":10651,"tokens_out":7495,"duration_ms":68147,"significance":"If the framework's claims were fully supported, the paper would address a real gap in coverage planning: semantic-aware speed allocation is genuinely underdeveloped, and replacing ESDF rebuilds with incremental KD-tree-based local optimization is a promising direction. The paper also ships an open-source implementation and includes real-robot validation, both of which are strengths. The significance is currently limited by the incorrect speed-law derivation, the unspecified safety-score function, and missing statistical detail in the empirical comparisons, so the headline state-of-the-art claim is not yet established.","major_comments":[{"comment":"The identity used to pass from Eq. (16) to Eq. (17) is not correct. For the kernel in Eq. (15), the integral over the disk ||p-p'||≤R is not 2f(R/σ)-1, where f is the standard Gaussian CDF. If G is intended to be a 2D-normalized Gaussian, the disk integral is 1 - exp(-R^2/(2σ^2)); with the kernel as written, which is 1D-normalized over two coordinates, the value is different again. Neither expression equals 2f(R/σ)-1. Because Eqs. (18) and (19) are obtained by inverting this integral, the dwell-time and speed allocation laws do not follow from the stated diffusion model. This is the load-bearing step for the coverage-uniformity claim, so the derivation must be corrected and the speed law re-derived.","section":"Section V, Eq. (17)"},{"comment":"Independently of the integration formula, the inversion conflates the dwell-time variable. In Eq. (16), the dwell time is t(p') and appears inside the integral over p'; in Eq. (17), it is factored out as t(p), evaluated at the query point p. The assumption of 'approximately uniform dwell time within the local Gaussian footprint' is not stated quantitatively and does not by itself justify replacing t(p') with t(p). The derivation needs either a formal argument for the factorization or a modified model before the closed-form inversion can be accepted.","section":"Section V, Eqs. (16)-(19)"},{"comment":"The safety score S_k in Eq. (22) is defined only as a weighted combination f(D, C_cont, C_feas), with no explicit form, weights, or normalization. Algorithm 1 and the sliding-window optimization both depend on comparing S_k with τ_safe, so the non-compliant segment identification is not reproducible and its properties cannot be checked. The threshold direction is also inconsistent: Algorithm 1 treats S_k < τ_safe as non-compliant, while the caption of Fig. 2 says segments whose score exceeds the threshold are gathered. Please specify f explicitly and reconcile the threshold convention.","section":"Section VI-B, Eq. (22)"},{"comment":"The empirical performance claims are presented as single point estimates without standard deviations, confidence intervals, or statistical tests. The text mentions '10 different cleaning tests' for one comparison, but Tables I and II do not report trial counts or variance. Additionally, no ablation isolates the RFICP speed allocation from the boustrophedon path layout and the local replanner, so the claimed improvements in cleaning uniformity cannot be attributed to the Gaussian-field speed law. Please add error bars, trial counts, significance tests, and an ablation of the speed-allocation component.","section":"Section VIII, Tables I and II"}],"minor_comments":[{"comment":"G is called a Gaussian kernel but is normalized as a one-dimensional density while being integrated over a two-dimensional area; please make the normalization convention explicit, for example by using 1/(2πσ^2) for a 2D Gaussian or by stating that the kernel is unnormalized.","section":"Section V, Eq. (15)"},{"comment":"The local cost terms J_obs, J_smooth, and J_len, as well as the B-spline basis N_{i,k} and its knot vector, are not defined; please provide explicit expressions so that the optimizer is reproducible.","section":"Section VI-C, Eqs. (23)-(24)"},{"comment":"Coverage completeness, cleaning uniformity, and the 'preset threshold' used to define cleaned area are described only qualitatively; please give precise definitions and the threshold value used in the experiments.","section":"Section VIII-B"},{"comment":"Several references use generic author names, for example 'J. Doe and J. Smith' in Ref. [14]; please verify these citations and replace any unverifiable entries with accurate bibliographic records.","section":"References [11], [12], [14]"}],"recommendation":"major_revision","confidential_remarks":"I recommend major revision rather than rejection because the empirical contributions may be salvageable with a corrected derivation and more complete reporting. The editorial office should verify the provenance of the cited literature, especially the entries with placeholder-like author names; if any reference is unverifiable, that is a separate integrity issue that must be resolved before publication. The system-integration aspects fit the journal's scope, but the theoretical section needs to meet the journal's standards before the paper can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe SHIFT planner is a real systems integration: semantic-aware speed modulation plus IKD-tree sliding-window local replanning, demonstrated on a drone in simulation and a vacuum robot in the lab. The local replanning story—IKD-tree distance queries instead of ESDF rebuilds, 1.6 ms and 0.9 ms per iteration versus 3–4 ms for EWOK and Fast-Planner—is the most credible part, and the comparison is useful even if not deeply controlled. The paper also earns credit for hardware experiments and an explicit open-source claim.\n\nThe load-bearing math, though, does not survive scrutiny. Equation (17) replaces the disk integral of the Gaussian kernel in Eq. (15) with 2f(R/σ)−1, a 1D CDF. The actual 2D disk integral is 1 − exp(−R^2/(2σ^2)), up to the normalization constant in Eq. (15), which is itself not 2D-normalized. So Eqs. (18) and (19), for dwell time and speed, do not follow from the stated diffusion model. There is also a second conflation: t(p′) under the integral in Eq. (16) becomes t(p) outside it, which is a heuristic, not a derivation. This means the uniform-coverage speed law—the part that makes RFICP novel—is unsupported as written. It might be salvageable as a heuristic or with a corrected derivation, but the paper does not provide either.\n\nOther soft spots line up with that. The safety score in Eq. (22) is never defined; f, Ccont, Cfeas, and the weights are all absent, so Algorithm 1 is not reproducible. Tables I and II report single numbers with no error bars or trial counts, and the 34.7% uniformity gain over baseline is presented without variance. The open-source pointer is a webpage URL, not a repository or commit hash, so the claim is hard to verify. The citation to [17] is too thin to support a nontrivial integral identity, and the related-work framing around Fast-Planner is loose.\n\nWho is this for? A robotics engineer looking for a practical semantic-coverage system will get a useful architecture and a plausible local-replanning recipe. A reader who checks derivations will stop at Section V. It deserves a serious referee because the system contribution and experiments are real, and the flaw is identifiable and fixable. I would send it to review with a clear instruction that Section V must be corrected or reframed before acceptance.","headline":"A genuinely useful systems integration with a central speed-law derivation that does not hold up; the local replanning results alone might justify a revised version, but Section V needs fixing before I would trust the coverage claims.","tokens_in":11128,"tokens_out":3762,"would_cite":false,"duration_ms":33772,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The SHIFT planner allocates dwell time from a Gaussian semantic field and refines only affected trajectory segments, yielding uniform coverage with millisecond replanning.","keywords":["coverage path planning","semantic mapping","Gaussian diffusion field","speed allocation","IKD-tree","sliding window optimization","dynamic obstacle avoidance","uniform coverage"],"falsifier":"Integrate the Gaussian kernel in Eq. (16) over the disk $\\|p-p'\\|\\le R$ numerically for several values of $R/\\sigma$ and compare the result with $2f(R/\\sigma)-1$; a mismatch means the dwell time and speed in Eqs. (18)-(19) do not solve the stated coverage equation. A second check is to simulate a single robot pausing along the proposed speed profile, accumulate the Gaussian kernel, and see whether the resulting coverage field matches $k A(p)-C_{\\rm target}$.","tokens_in":10048,"feed_emoji":"🧹","tokens_out":8196,"duration_ms":67493,"temperature":0.7,"pith_summary":"The paper proposes SHIFT, a planning framework intended to make robotic coverage uniform, terrain-adaptive, and inexpensive to replan when obstacles move. It claims that a Gaussian diffusion field over semantic attributes like dirtiness and dryness can set the robot's dwell time, so areas that need more attention receive proportionally more cleaning effort, and that an incremental KD-tree plus sliding-window optimization can handle dynamic obstacles by refining only the affected trajectory segments. The authors validate the framework on a simulated agricultural drone and a physical robot vacuum, reporting higher coverage completeness, higher cleaning uniformity, and lower per-iteration planning time than uniform-speed and ESDF-based baselines.","feed_headline":"Speed follows a dirtiness field; obstacle replanning takes ~1 ms","feed_subtitle":"It paces coverage by semantic importance and refines only affected path segments, cutting replanning to under 2 ms.","key_machinery":"The load-bearing object is the Gaussian coverage kernel of Eq. (15), which converts a pause of duration $t$ at the robot position into coverage effort at every point within radius $R$, and the resulting closed-form speed law $v(p)\\approx-\\lambda/\\ln(1-(k A(p)-C_{\\rm target})/(2f(R/\\sigma)-1))$. The second mechanism is the IKD-tree distance field, which supplies near-lossless obstacle-distance queries for the A* initial guess and for safety-region scoring, and the adaptive sliding-window optimizer that refines only non-compliant segments and reconnects them with B-splines. Together they separate global coverage allocation from local reactive replanning, which is what keeps the per-iteration cost low.","core_discovery":"The central claim is that coverage quality is governed by a field model: the robot's action spreads as a Gaussian kernel $G(p;p')$ over a disk of radius $R$, and the dwell time $t(p)$ at each waypoint is chosen so the accumulated coverage equals the semantic demand $k A(p)-C_{\\rm target}$, giving the speed law $v(p)\\approx -\\lambda/\\ln\\bigl(1-(k A(p)-C_{\\rm target})/(2f(R/\\sigma)-1)\\bigr)$ (Eq. 19). On top of this speed allocation, the trajectory is laid out as boustrophedon waypoints on a curvature-filtered elevation surface and then locally re-planned by IKD-SWOpt, which uses IKD-tree distance queries, an A* search with a clearance penalty, and a GPU-parallel sliding-window optimizer followed by B-spline reconnection. In the reported benchmarks the framework reaches 98.6% coverage completeness and 95.2% cleaning uniformity on a vacuum-cleaner test, and 0.9 ms per-iteration planning time with GPU acceleration in obstacle-dense 3D maps.","pith_inferences":["If Eq. (17)'s disk integral is replaced by the exact value $1-\\exp(-R^2/(2\\sigma^2))$, the constant in the speed law changes but the monotone dependence on $A(p)$ survives, so the qualitative 'slow down where it is dirtier' behavior would likely persist; the quantitative uniformity guarantee would need rederivation.","The semantic field $A(p)$ could be estimated online from cleaning feedback, turning the planner into a closed-loop coverage controller that re-paces itself as it senses remaining dirt.","The same field-and-dwell-time derivation, with the kernel integral computed for the actual footprint shape, could serve as a general coverage-cost model for spraying, painting, or disinfection tasks where action effect spreads spatially."],"forward_implications":["Semantic-field-paced coverage can replace uniform-speed profiles without sacrificing path efficiency: the reported overlap rate is 4.3%, against 10.2% for the uniform-speed baseline.","Local replanning no longer requires rebuilding an ESDF over the whole map; only non-compliant segments are re-optimized, giving 1.6 ms per iteration on CPU and 0.9 ms on GPU.","The same speed-allocation law transfers across platforms and semantic labels, from dryness-triggered irrigation by a drone to dirtiness-triggered cleaning by a vacuum.","GPU-parallel segment optimization makes the sliding-window refinement scalable to denser obstacle fields without changing the global coverage plan."],"supporting_citations":[{"why":"Supplies the Gaussian CDF identity used to convert the disk integral in Eq. (17) into the dwell-time formula.","marker":"[17]"},{"why":"Provides the incremental KD-tree whose distance queries drive the A* heuristic and safety scoring in IKD-SWOpt.","marker":"[18]"},{"why":"Provides the L-BFGS solver used to minimize the local trajectory cost in the sliding-window optimization.","marker":"[19]"},{"why":"Fast-Planner is the ESDF-based kinodynamic baseline compared against IKD-SWOpt for local replanning time and path quality.","marker":"[23]"},{"why":"EWOK is the ESDF-based B-spline local planner used as the second replanning baseline.","marker":"[24]"},{"why":"Serves as the coverage-path-planning baseline in the cleaning-uniformity benchmark.","marker":"[20]"},{"why":"Serves as the K-Means+DRL coverage baseline in the cleaning-uniformity benchmark.","marker":"[21]"},{"why":"Establishes the boustrophedon cellular decomposition pattern used to sequence the terrain waypoints.","marker":"[7]"}],"fun_headline_variants":["Gaussian dirtiness field sets speed; replanning in <1 ms","Field-modulated speed: 98.6% coverage, 95.2% uniformity","Adaptive coverage: speed by dirtiness, replanning <1 ms","SHIFT Planner: field-driven speed, sub-ms replanning"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a two-dimensional Gaussian spread of cleaning effort over a circular footprint can be treated as a one-dimensional Gaussian cumulative probability; if that simplification is not faithful to the stated two-dimensional model, the derived speed profile does not actually implement the claimed uniform-coverage law.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian dirtiness field sets speed; replanning in <1 ms","Field-modulated speed: 98.6% coverage, 95.2% uniformity","Adaptive coverage: speed by dirtiness, replanning <1 ms","SHIFT Planner: field-driven speed, sub-ms replanning"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001128,"raw_usage":{"total_tokens":4679,"prompt_tokens":921,"completion_tokens":3758,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":3673}},"tokens_in":537,"tokens_out":3758,"duration_ms":25272,"temperature":1.0,"reasoning_tokens":3673,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T15:41:20.669465+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Integrate the Gaussian kernel in Eq. (16) over the disk $\\|p-p'\\|\\le R$ numerically for several values of $R/\\sigma$ and compare the result with $2f(R/\\sigma)-1$; a mismatch means the dwell time and speed in Eqs. (18)-(19) do not solve the stated coverage equation. A second check is to simulate a single robot pausing along the proposed speed profile, accumulate the Gaussian kernel, and see whether the resulting coverage field matches $k A(p)-C_{\\rm target}$.","supporting_citations":[{"cited_title":"Pishro-Nik, Introduction to probability, statistics, and random processes","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian CDF identity used to convert the disk integral in Eq. (17) into the dwell-time formula."},{"cited_title":"Fast-planner: An efficient coverage path planning framework for unmanned aerial vehicles,","cited_arxiv_id":null,"evidence_quote":"Fast-Planner is the ESDF-based kinodynamic baseline compared against IKD-SWOpt for local replanning time and path quality."},{"cited_title":"A clustering-based coverage path planning method for autonomous heterogeneous uavs,","cited_arxiv_id":null,"evidence_quote":"Serves as the coverage-path-planning baseline in the cleaning-uniformity benchmark."},{"cited_title":"Cooperative coverage path planning for multi-mobile robots based on improved k-means clustering and deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Serves as the K-Means+DRL coverage baseline in the cleaning-uniformity benchmark."},{"cited_title":"Coverage path planning: The boustrophedon cellular decomposition,","cited_arxiv_id":null,"evidence_quote":"Establishes the boustrophedon cellular decomposition pattern used to sequence the terrain waypoints."}],"review_version":1}