{"id":"a11dfc83-89e4-47c9-9e01-4892985b9888","arxiv_id":"2507.14099","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A PRM-based planner with cached 'highway' paths and a hand-crafted probabilistic selection rule claims faster multi-goal planning for underwater manipulation.","lead":"This paper describes an underwater manipulator planner that stores previously computed paths as 'highway nodes' on top of a probabilistic roadmap, then selects among them with a probability formula. It reports lower planning times than plain PRM plus A* in simulation, and similar trajectory accuracy on a physical BlueROV2 with a Reach Alpha 5 arm.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's cached-path splice is not a continuous path: HMS stores a full path keyed by its destination, so PartialA ∥ HMS[oNode].path ∥ PartialB starts at the old origin, not at currentNode. The claimed speedup rests on this undefined reuse.","rationale":"The reader identified the selection probability P(Mi | G, E) and the unspecified Bayesian network as the weakest assumptions. I find a more elementary and more load-bearing problem earlier in the same algorithm: the cached-path representation and its concatenation are not defined in a way that yields a valid path from the current node to the goal. The only cache-write in Algorithm 1 stores a full path keyed by its destination, so retrieving that path and splicing it between two A* segments is not a continuous operation unless an unstated re-rooting or highway-segment representation exists. Even if that ambiguity were resolved, the complexity argument in Section III-D has no proof that the two partial A* searches have depth d' ≪ d; line 8 can span the whole roadmap when the selected oNode is close to the goal but far from the current configuration. The paper does present real tank experiments and a reproducible-looking setup, which is credit in its favor, but those experiments do not report planning times, and the simulation figure lacks error bars. The concern is addressable: the authors need to specify the exact data structure for HMS[oNode].path, show how splices maintain continuity, and verify the d' bound empirically. That is why I keep a conditional disposition rather than rejecting outright, but my reason is different from the reader's and is more fundamental.","tokens_in":9899,"tokens_out":8080,"duration_ms":563652,"concrete_test":"Implement Algorithm 1 literally on the Fig. 4 simulator for two consecutive goals, s→g1→g2, and instrument the output: (i) assert Π_2[0]=g1 and Π_2[-1]=g2; (ii) assert every consecutive waypoint pair is connected by an edge or valid local path in the PRM; (iii) if g1 is selected as oNode for g2, print the waypoint sequence of PartialA ∥ HMS[g1].path ∥ PartialB. The literal algorithm will either fail these assertions or require an extra re-rooting step that is absent from Algorithm 1. Repeat over 100 random 10-goal sequences and report the discontinuity rate; if it is nonzero, the experiments in Fig. 4 cannot have been produced by Algorithm 1 as written.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central speedup claim depends on Algorithm 1's path-reuse step, but as written that step is not a valid path operation. Line 21 caches the entire planned path under the new goal: HMS[gi] ← (Πi, 1.0). On a later query, line 7 retrieves HMS[oNode].path; since the key is the destination, this is a path from whatever node was current at caching time to oNode. The output at line 11 concatenates PartialA (currentNode → oNode), that cached path, and PartialB (oNode → gi). For this concatenation to be a route from currentNode to gi, the cached path would have to start at oNode and end at the start of PartialB. Instead it starts at the old origin and ends at oNode, producing a backtrack and a discontinuity. No alternative semantics for HMS[oNode].path are defined, and no continuity/splice condition is given. Separately, Section III-D's claimed ratio TA*/THMS ≈ b^d / b^{d'} with d' ≪ d is not supported by the algorithm: line 8, A*(E, currentNode, oNode), can still require a full-depth search when the only cached node near the goal is far from the current configuration, so d' is not bounded. The tank section also states that speed gains are 'not explicitly shown here but implied.' The validity of the whole efficiency claim therefore currently rests on an undefined path-reuse representation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes the Adaptive Heuristic Motion Planner (AHMP), which layers a Heuristic Motion Space (HMS) on top of a dense PRM roadmap to cache and reuse previously computed 'highway' paths for repeated or multi-goal motion planning in underwater manipulation tasks. A Bayesian Network is said to update uncertainty estimates from sensor data and to bias the selection of cached paths. The central claims are that AHMP reduces the effective A* search depth from d to d' << d, yielding the approximate speedup ratio T_A*/T_HMS ≈ b^d / b^{d'} in Section III-D, and that it produces trajectories of accuracy comparable to plain PRM while being more scalable. The paper reports simulation comparisons against PRM+A* and RRT and physical tank experiments with a BlueROV2 and Reach Alpha 5 manipulator.","tokens_in":10270,"tokens_out":2852,"duration_ms":38841,"significance":"The idea of caching and reusing high-value motion segments for repeated manipulation tasks is practically relevant for underwater inspection and welding scenarios, and the paper includes a real-hardware demonstration, which is a strength. However, the central efficiency claim currently rests on an invalid path-splicing operation in Algorithm 1 and on an unverified depth-reduction argument in Section III-D. The Bayesian Network component, which is central to the claimed adaptability, is specified only at the level of a generic factorization, with no structure, parameters, or update rules given. If these load-bearing gaps were repaired, the approach could be a useful contribution, but as written the manuscript does not establish the claimed speedup or the role of the probabilistic components.","major_comments":[{"comment":"The cached-path splice is not a valid path operation as written. Line 21 caches the full planned path under the goal key: HMS[gi] ← (Πi, 1.0). On a later query, line 7 retrieves HMS[oNode].path, which is therefore a path from whatever node was current at caching time to oNode, not a path from oNode to the start of PartialB. The concatenation PartialA ∥ HMS[oNode].path ∥ PartialB in line 11 is therefore discontinuous and may contain backtracking. The paper never defines an alternative semantics for HMS[oNode].path or provides a splice/continuity condition. Since the entire speedup claim depends on reusing this path, this is a load-bearing flaw that must be fixed by specifying the cached path representation (e.g., storing paths by origin and destination) and by proving that the concatenation is a valid continuous path in the current environment.","section":"Algorithm 1, lines 7-11"},{"comment":"The claimed time ratio T_A*/T_HMS ≈ b^d / b^{d'} is not established by the algorithm. The reuse of a cached path is the mechanism, but line 8 still requires A*(E, currentNode, oNode), and line 9 requires A*(E, oNode, gi). If oNode is selected only by proximity to the goal, the segment from currentNode to oNode may require a full-depth search of the roadmap, so d' is not bounded by construction. Moreover, the ratio as stated is largely a restatement of the caching design rather than an independently verified empirical result. The paper should either provide a formal bound on d' under the selection criterion or report measured search-node expansions and wall-clock times across the full goal sequence, including the cost of populating the HMS.","section":"Section III-D, Eq. (ratio)"},{"comment":"The selection model is underspecified. The formula P(Mi | G, E) ∝ exp(−λ Ui) / (1 + d(Mi, G)) uses Ui, λ, and d(Mi, G), but Ui is never defined operationally, λ is never set or analyzed, and no sensitivity or ablation is given. The Bayesian Network is described only through the generic factorization P(X1,...,Xn) = ∏ P(Xi | Pa(Xi)); the DAG structure, the random variables, the conditional probability tables, and the update rule from sensor data are never specified. Because the paper claims that BN updates refine path success probabilities and guide real-time adaptation, this is load-bearing. The authors should fully define the BN and provide an ablation comparing the proposed selection with distance-only and random selection to show that the probabilistic ranking is what provides the benefit.","section":"Section III-C, selection probability and BN"},{"comment":"The paper repeatedly states that paths are optimized by minimizing a composite cost function accounting for distance, uncertainty, energy consumption, and execution time (Abstract and Section I), but no such cost function is defined anywhere in the Methodology. Algorithm 1 calls A* without specifying an edge-cost or heuristic function, and Section III-C only describes a probability proxy for selecting cached nodes, not a path cost. This mismatch makes it impossible to reproduce the planner or interpret the reported runtimes. Either define the composite cost and its use in A*, or remove this claim and state the actual objective.","section":"Abstract and Introduction, composite cost function"},{"comment":"The experimental evidence for the real-time efficiency claim is incomplete. Figure 4 shows mean execution times over 5 runs but no error bars or per-run variation, so the claimed stability of AHMP is not statistically supported. In the tank experiments, the speed gains are explicitly deferred: the text says 'not explicitly shown here but implied by the use of HMS' (Section IV-B). Table I reports 'Mean Abs. Error Across Joints' but does not clearly define the error: whether it is the difference between planned and executed joint trajectories, or between the PRM and AHMP paths. These omissions matter because the accuracy result is one of the two central claims. Please report distributions, define the error metric, and provide actual runtime measurements from the hardware experiments.","section":"Section IV and Table I"}],"minor_comments":[{"comment":"Typo: 'With reward to runtimes' should be 'With regard to runtimes'.","section":"Section III-D"},{"comment":"The parameter 'max iter rrt' is mentioned but never defined, and the figure legend does not state the RRT iteration counts corresponding to the four colors.","section":"Section IV-A"},{"comment":"The related-work discussion references several memory-based planners, including Motion Memory and EB-RRT, but the experimental comparison includes only PRM+A* and RRT; a direct comparison with at least one of these memory-based planners would strengthen the claims about the contribution of HMS.","section":"Section II"},{"comment":"In Table I, the standard deviation columns show large run-to-run variation relative to the mean errors; please clarify whether the table reports errors across joints, across goals, or across trials, and correct the caption accordingly.","section":"Section IV-B"},{"comment":"The notation is inconsistent: the algorithm uses 'BN' as an input to the heuristic selection, while the text describes a probabilistic evaluation P(u | gi, BN); the relation between the BN output and the heuristic function heuristic(u, gi) is never defined.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a relevant problem and includes a real-hardware demonstration, which is a positive element. However, the central algorithm as presented contains an invalid path-concatenation step, and the theoretical speedup argument is not backed by a formal bound or by sufficient empirical data. The Bayesian Network component is essentially unspecified, and the experiments lack error bars and a defined error metric. These are substantial but fixable issues; the paper should be returned for major revision rather than rejected outright. I would also encourage the editor to ask for a clear statement of the exact path representation stored in the HMS and for measured runtime data from the tank experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about this paper. First, it has a real, demonstrable bug in the central algorithm: the path-reuse step in Algorithm 1 concatenates a cached path that starts at an old origin, not at the current node, so the output is not a continuous route. That is load-bearing, because the entire efficiency claim rests on this splice. Second, the Bayesian network described in the abstract and architecture is never actually defined—no structure, no CPTs, no update equations—so the 'adaptive' half of the method is not reproducible.\n\nWhat is new: the specific combination of a PRM with a cached-highway memory and a hand-crafted selection rule for an underwater vehicle-manipulator system. The simulation shows a clear scaling advantage at high sample counts (2.2s vs 5.6s at 30k samples), and the tank experiments demonstrate that the planned trajectories closely match baseline PRM. The authors are honest about the fact that speed gains in the tank are implied, not measured.\n\nThe soft spots beyond the algorithm bug: the complexity ratio in Section III-D is largely a restatement of the caching idea—of course reusing paths avoids searching, so the ratio is not an independent discovery. The composite cost function promised in the abstract never appears. Fig. 4 lacks error bars despite averaging five runs. Most importantly, the two closest experience-based baselines (Motion Memory, EB-RRT) are cited but never compared, so the incremental value is unclear.\n\nThat said, the core caching idea is defensible, the paper is clearly written, and all the missing pieces are addressable in a revision. This is not a desk-reject; it deserves serious refereeing. A referee should ask for: a corrected algorithm with a clear semantics for cached path reuse, either a real BN or removal of that claim, error bars, and those baseline comparisons. I would not accept it as is.\n\nWho it is for: roboticists working on practical manipulation planning, especially underwater, who want a lightweight experience-based planner. I'd bring it to a reading group to dissect the algorithm, but I probably wouldn't cite it in my own work until the details are fixed.","headline":"A sensible caching-on-PRM idea that is undercut by an invalid path-splice in Algorithm 1 and a Bayesian network that never materializes.","tokens_in":10728,"tokens_out":5481,"would_cite":false,"duration_ms":57111,"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":"Caching and reusing 'highway' paths on a dense roadmap cuts the exponential cost of repeated underwater motion planning while keeping trajectory fidelity close to baseline PRM.","keywords":["underwater manipulation","motion planning","Probabilistic Roadmap (PRM)","Heuristic Motion Space","Bayesian network","experience reuse","A* search","underwater vehicle-manipulator system"],"falsifier":"Ablate the learned selection: replace the scoring rule with uniform-random choice of a cached node within the same distance threshold, keeping the same stitch-and-fallback logic. If random caching matches the reported speedups and the joint-error fidelity of Table I in the same tank scenario, then the ranking—and the Bayesian updates behind it—are not what carries the result. Separately, instrument the planner to measure effective A* expansion depth $d'$ with and without the HMS on identical goal sets; the claimed time ratio $b^d/b^{d'}$ makes a quantitative prediction that can be checked directly.","tokens_in":9686,"feed_emoji":"🤖","tokens_out":15114,"duration_ms":141348,"temperature":0.7,"pith_summary":"The paper sets out to break the usual trade-off in Probabilistic Roadmap (PRM) planning: dense sampling is needed to thread cluttered underwater workspaces, but dense roadmaps make each A* query exponentially expensive. Its proposal, the Adaptive Heuristic Motion Planner (AHMP), keeps the dense roadmap and layers a Heuristic Motion Space (HMS) on top of it—a memory of 'highway' nodes and previously successful paths. For each new goal, the planner selects a cached highway approach node, runs two short A* searches from the current state to that node and from the node to the goal, and splices the cached path between them; only if a segment fails does it fall back to full A*. The claimed payoff is that per-goal cost drops from $O(b^d)$ to roughly $O(b^{d'})$ with $d'\\ll d$, so multi-goal tasks like underwater welding or inspection stay real-time even at 30,000 samples, while tank experiments on a vehicle–manipulator system show joint trajectories within about 0.28–0.39 rad of baseline PRM. A Bayesian Network is meant to feed uncertainty estimates from live sensor data into the choice of cached path, giving the planner a way to favor routes that are not just short but likely to succeed.","feed_headline":"Reusing cached paths keeps underwater replanning fast as goals grow","feed_subtitle":"At 30,000 roadmap samples the cached-path planner takes 2.2 s where plain A* needs 5.6 s.","key_machinery":"The carrying mechanism is the Heuristic Motion Space (HMS)—a cache of highway nodes and the paths connecting them, layered on top of the base PRM rather than replacing it. Each stored motion primitive $M_i$ carries an uncertainty estimate $U_i$; selection on a query is driven by the score $\\exp(-\\lambda U_i)/(1+d(M_i,G))$ with a distance threshold $\\tau$ (Algorithm 1), and the planner executes A* only on the two residual segments $(current \\to oNode)$ and $(oNode \\to goal)$. The exponential speedup claim rests on the depth reduction: the cached highway replaces the deepest part of the search, so effective A* depth falls from $d$ to $d'\\ll d$ and the time ratio $T_{\\mathrm{A}^*}/T_{\\mathrm{HMS}}\\approx b^d/b^{d'}$ follows. A Bayesian Network—a directed acyclic graph over environment and robot variables, with the stated factorization $P(X_1,\\dots,X_n)=\\prod_i P(X_i \\mid \\mathrm{Pa}(X_i))$—is included to update the uncertainty terms from real-time sensor data, and after each successful plan the stored probabilities are reweighted by $p_{\\mathrm{new}}(u)\\propto p_{\\mathrm{old}}(u)\\exp(-\\alpha\\|\\Pi_i\\|)$ and renormalized. The system also caches every new path under the goal at weight 1.0, so the memory grows with mission experience.","core_discovery":"Stated on the paper's own terms, the discovery is that memory can substitute for search without sacrificing coverage. The AHMP deliberately builds a dense PRM—up to tens of thousands of collision-free samples—so that narrow passages and fine obstacle clearance are represented, then overlays a Heuristic Motion Space that records frequently traversed sub-paths (motion primitives) and the highway nodes they connect. On a query, the planner scores each cached primitive by $P(M_i \\mid G, E) \\propto \\exp(-\\lambda U_i)/(1 + d(M_i, G))$, where $U_i$ is the primitive's uncertainty estimate and $d(M_i, G)$ its distance to the goal in configuration space; the best-scoring node within a threshold $\\tau$ becomes the approach node, and the final trajectory is the concatenation of two short A* segments with the cached highway path. The paper's complexity argument is that this replaces one deep search per goal with two shallow ones: total cost goes from $N \\cdot O(b^d)$ to $N \\cdot O(|HMS|) + \\sum_i O(b^{d'_i})$ with $d'_i \\ll d$, so the time ratio $T_{\\mathrm{A}^*}/T_{\\mathrm{HMS}} \\approx b^d / b^{d'}$ grows exponentially in the depth saved. Experiments in a tank with a BlueROV2 and Reach Alpha 5 arm show the AHMP tracking baseline PRM joint trajectories with mean absolute errors around 0.28–0.39 rad across five test motions, while RRT deviates more; in simulation, AHMP runtime stays stable as sample count and goal count grow, ending near 2.2 s where plain PRM+A* reaches about 5.6 s at 30,000 samples.","pith_inferences":["A direct consequence the paper does not state: the speedup should be largest for goals near previously cached highways and negligible for the first query in an empty cache, so missions could deliberately warm up the HMS with bootstrapping goals to maximize the gain.","A natural experiment the paper leaves open: freeze the Bayesian updates and compare cached-path selection against the live-BN version on the same sensor stream, to isolate how much the Bayesian Network contributes beyond the geometric score.","The same highway-caching pattern could transfer beyond underwater systems to any manipulator with recurring goal sets—warehouse arms, inspection robots—where the configuration space is cluttered and the same sub-paths are traversed repeatedly.","The validation is kinematic, with the vehicle stabilized so arm motion is decoupled from the base; extending the planner to whole-body vehicle–manipulator coordination during grasping would test whether cached paths remain safe when the base moves."],"forward_implications":["Multi-goal missions in cluttered environments—underwater spot welding, inspection, cleaning—can keep real-time planning at roadmap densities where naive PRM+A* becomes impractical: at 30,000 samples the reported runtime is about 2.2 s versus 5.6 s for plain A*.","The speedup does not come from thinning the roadmap, so dense-sample safety clearance and narrow-passage connectivity are preserved; trajectory accuracy stays close to baseline PRM (mean joint error roughly 0.28–0.39 rad in the tank tests).","When no cached highway is usable, the method degrades gracefully to full A*, so a cold start or a changed obstacle layout does not leave the planner without a path.","Because the HMS accumulates experience online with no training phase, the planner can carry its memory into unexplored environments: each executed plan enriches the cache that speeds up later queries."],"supporting_citations":[{"why":"Foundational Probabilistic Roadmap paper: supplies the sampling-based roadmap that AHMP keeps dense and searches over with A*.","marker":"[18]"},{"why":"Prior experience-based planner (EB-RRT) that reuses stored experience graphs; AHMP is positioned against its lack of real-time adaptability and uncertainty handling.","marker":"[13]"},{"why":"Motion Memory, the closest prior work on reusing past trajectories to bias sampling; the HMS extends this idea with a highway cache and probabilistic reasoning.","marker":"[19]"},{"why":"Multi-heuristic A* approach demonstrating existing heuristic-combination methods, which the paper contrasts with its memory-based, uncertainty-aware selection.","marker":"[25]"},{"why":"BlueROV2 platform reference: the underwater vehicle used in the tank experiments that validate the planner's trajectories.","marker":"[29]"},{"why":"Reach Alpha 5 manipulator reference: the arm whose joint trajectories are planned, executed, and compared in the experiments.","marker":"[30]"}],"fun_headline_variants":["Cached paths slash underwater replanning time","Heuristic memory replaces deep search in water","Bayesian-cached motion beats PRM+A* in tests","Memory-driven planning speeds underwater manipulation","Substitute memory for search in underwater tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speedup rests on the unverified assumption that the hand-crafted selection score $\\exp(-\\lambda U_i)/(1+d(M_i,G))$ with threshold $\\tau$ reliably picks cached highway paths whose stitched segments match full A* quality—and that the Bayesian Network, whose structure and update rules are never specified, supplies meaningful uncertainty values $U_i$ from sensor data.","fun_headline_variants_meta":{"raw":{"variants":["Cached paths slash underwater replanning time","Heuristic memory replaces deep search in water","Bayesian-cached motion beats PRM+A* in tests","Memory-driven planning speeds underwater manipulation","Substitute memory for search in underwater tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000177,"raw_usage":{"total_tokens":1377,"prompt_tokens":1114,"completion_tokens":263,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":730,"completion_tokens_details":{"reasoning_tokens":194}},"tokens_in":730,"tokens_out":263,"duration_ms":3720,"temperature":1.0,"reasoning_tokens":194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:01:03.883224+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ablate the learned selection: replace the scoring rule with uniform-random choice of a cached node within the same distance threshold, keeping the same stitch-and-fallback logic. If random caching matches the reported speedups and the joint-error fidelity of Table I in the same tank scenario, then the ranking—and the Bayesian updates behind it—are not what carries the result. Separately, instrument the planner to measure effective A* expansion depth $d'$ with and without the HMS on identical goal sets; the claimed time ratio $b^d/b^{d'}$ makes a quantitative prediction that can be checked directly.","supporting_citations":[{"cited_title":"Probabilistic roadmaps for path planning in high- dimensional configuration spaces,","cited_arxiv_id":null,"evidence_quote":"Foundational Probabilistic Roadmap paper: supplies the sampling-based roadmap that AHMP keeps dense and searches over with A*."},{"cited_title":"Repetitive path planning with experience-based bidirectional rrt,","cited_arxiv_id":null,"evidence_quote":"Prior experience-based planner (EB-RRT) that reuses stored experience graphs; AHMP is positioned against its lack of real-time adaptability and uncertainty handling."},{"cited_title":"Motion memory: Leveraging past experiences to accelerate future motion planning,","cited_arxiv_id":null,"evidence_quote":"Motion Memory, the closest prior work on reusing past trajectories to bias sampling; the HMS extends this idea with a highway cache and probabilistic reasoning."},{"cited_title":"Multirepresentation, mul- tiheuristic a* search-based motion planning for a free- floating underwater vehicle-manipulator system in un- known environment,","cited_arxiv_id":null,"evidence_quote":"Multi-heuristic A* approach demonstrating existing heuristic-combination methods, which the paper contrasts with its memory-based, uncertainty-aware selection."},{"cited_title":"Blue robotics – high-quality marine robotics components,","cited_arxiv_id":null,"evidence_quote":"BlueROV2 platform reference: the underwater vehicle used in the tank experiments that validate the planner's trajectories."},{"cited_title":"Reach robotics – underwater manipula- tors,","cited_arxiv_id":null,"evidence_quote":"Reach Alpha 5 manipulator reference: the arm whose joint trajectories are planned, executed, and compared in the experiments."}],"review_version":1}