{"id":"19c4a467-3407-4d45-bc4a-052f9f45142d","arxiv_id":"2608.12763","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"ARIES-Mission2 couples zero-shot vision-language target detection with a metaheuristic traveling-salesperson solver and reports shorter UAV mission routes than manual human planning on the UAV-VLPA-nano-30 benchmark.","lead":"This paper presents ARIES-Mission2, a pipeline that combines a vision-language model for finding targets in satellite images with a traveling-salesperson solver based on particle swarm optimization to plan short drone routes. On the UAV-VLPA-nano-30 benchmark it reports a total route 9.5% shorter than a human expert's plan and runs about 3.6 times faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline distance comparisons are unverified because target-set equivalence is never established: no per-task waypoint counts or detection precision/recall are reported, so the 9.5% margin over HumanPlan may reflect fewer/different targets rather than better routing.","rationale":"The reader's weakest assumption points to identical target sets, and this stress test confirms that as the most load-bearing issue. I checked the arithmetic in Table I and the totals are internally consistent, so the concern is not about calculation errors. The missing piece is empirical comparability: distance-only tables cannot distinguish better routing from fewer targets. No detection metrics or waypoint counts appear anywhere in the paper, despite the benchmark's stated requirement to extract target points. Because the central claim is a head-to-head distance comparison, this gap directly undermines the conclusion. I considered whether the lack of code or prompts is a stronger issue, but that is a reproducibility concern rather than a correctness threat to the reported numbers. The reader's conditional verdict already captures the right level of confidence, so I recommend no change.","tokens_in":9027,"tokens_out":6793,"duration_ms":55143,"concrete_test":"For each of the 30 tasks, compute Molmo detection precision/recall against the benchmark's ground-truth target sets (or a fixed reference set) and report the number of waypoints used by ARIES-Mission2, the raw VLA baseline, and HumanPlan. Then rerun all four route planners on the identical ground-truth waypoint set per task and recompute the Table I totals. If 62.43 km is no longer below 69.00 km under identical target sets, the headline claim is an artifact of target-set mismatch; if it remains below, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III.A defines the benchmark as requiring target-point extraction plus closed-loop trajectory planning, and Table I reports only distances. Section III.B's only detection evidence is a qualitative claim for Task 6. The paper never reports detection precision, recall, per-task target counts, or an explicit statement that ARIES-Mission2, the raw VLA baseline, and HumanPlan all plan over exactly the same waypoint set. If the VLM misses or hallucinates targets, the closed-tour length can change independently of routing quality, because removing waypoints from a metric tour cannot increase the optimal tour length. Thus the 21.6% and 9.5% reductions could be artifacts of unequal target sets, and the central claim that decoupling perception from TSP optimization beats human planning is not falsifiably supported. The HumanPlan protocol is also unspecified: no details on the interface, instructions, or whether the human used the same imagery and target definitions.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ARIES-Mission2, a zero-shot Vision-Language-Action (VLA) framework that decouples visual-semantic perception from physical route optimization. The front end uses DeepSeek-V3 for task parsing and Molmo-7B for zero-shot target grounding, converting detected pixel locations to GPS waypoints via geospatial interpolation. The back end formulates multi-target UAV traversal as a Traveling Salesperson Problem (TSP) and compares four candidate routes—the raw VLM visiting order and routes optimized by PSO, GPSO, and IPSO—selecting the minimum-cost closed-loop route. Experiments on the UAV-VLPA-nano-30 benchmark report a total flight distance of 62.43 km, a 21.6% reduction versus the unoptimized VLA baseline (79.66 km) and a 9.5% reduction versus manual human planning (69.00 km). The complete 30-task workflow takes 575.40 s, about 3.6 times faster than human expert planning. The paper also reports component-level timing and a scalability analysis indicating that the TSP solver's computation time grows more slowly than the VLM's as the number of targets increases.","tokens_in":9230,"tokens_out":4170,"duration_ms":38803,"significance":"If the quantitative claims are validated, the decoupled architecture is a practically interesting and lightweight contribution: it shows that a small metaheuristic TSP solver can repair the poor visiting orders produced by a zero-shot VLM, without retraining or fine-tuning. The pipeline is transparent and the components are standard, making the framework easy to reproduce. The paper also provides a useful decomposition of where latency actually resides (VLM inference dominates by two orders of magnitude). However, the central route-quality claims are not yet falsifiably supported because the paper does not establish that all compared planners used identical target waypoint sets, and the reported totals are best-of-four selections from single stochastic runs. These gaps are load-bearing for the headline conclusions, so the contribution is currently conditional.","major_comments":[{"comment":"The headline route-length comparisons require that all compared planners visit exactly the same set of target waypoints, but the paper never establishes this. Section III.A defines the benchmark as requiring target-point extraction plus closed-loop trajectory planning, and Table I reports only route distances; there is no per-task count of detected waypoints, no detection precision or recall, and no explicit statement that ARIES-Mission2, ARIES-Mission2 w/o MAs, and HumanPlan all planned over identical target sets. The only detection evidence is the qualitative claim in Section III.B referencing Fig. 6. Because removing waypoints from a closed metric tour cannot increase the optimal tour length, the reported 21.6% and 9.5% reductions could be artifacts of visiting fewer or different points rather than of better routing. Please report per-task waypoint counts and detection metrics, and either enforce or explicitly verify identical target sets across all planners.","section":"Section III.A, Table I"},{"comment":"The total distance of 62.43 km is the sum of the per-task minimum over the four candidate routes (Raw, PSO, GPSO, IPSO), i.e., a best-of-four selection. Since PSO, GPSO, and IPSO are stochastic optimizers and the paper reports a single run with no variance information, this headline number is an apparent minimum selection and may be an optimistic realization; the internal consistency of the numbers does not establish statistical robustness. The hyper-parameters (Pop=30, Iter_max=150) are fixed without sensitivity analysis. Please report multiple independent runs (e.g., 10–30 seeds) with means and standard deviations, and ideally compare against an exact TSP solver or a bound to quantify the optimality gap, as well as against the original UAV-VLA system (reference [17]).","section":"Section II.F, Eq. (16), Table I"},{"comment":"The HumanPlan baseline is under-specified. The paper does not state how many human experts participated, what interface or instructions they were given, whether they used the same satellite imagery and target definitions as the automated pipeline, or whether the 2100 s in Table II represents a single expert's time for all 30 tasks. Without this protocol information, the claims of 9.5% shorter distance and 3.6 times faster than human expert planning are not robust. Please specify the planning protocol, number of experts, and how target sets were fixed for the human planners.","section":"Section III.B, Table II"}],"minor_comments":[{"comment":"The Tent chaotic map formula is ambiguous: the '+ U(-0.001,0.001)' term appears to apply to only the second branch, and the initialization of X_{0,j} is not defined. Please clarify.","section":"Section II.E, Eq. (12)"},{"comment":"The position variable X_i^{(t+1)} is used on both sides of the update, which is an overloading of notation; consider using a temporary variable for the intermediate position.","section":"Section II.E, Eq. (15)"},{"comment":"The text says the minimum target count of 4 is used as a baseline, but the horizontal axis starts at 6 targets; the figure also lacks error bars and a description of how the computation ratio was averaged. Please clarify.","section":"Figure 10 and surrounding text"},{"comment":"The table is formatted in a run-on manner in the text; please reformat with clear per-task rows and columns, and add a column for target counts if available.","section":"Table I"},{"comment":"The phrase 'breakthrough flight efficiency' in the contribution list is promotional; consider a more neutral formulation.","section":"Section I"}],"recommendation":"major_revision","confidential_remarks":"The reference list contains a large share of self-citations by the corresponding author and co-authors (e.g., refs [4]–[16], [18]–[26]), several of which are arXiv preprints with future dates. While this does not affect the technical evaluation, the editorial board may wish to check for citation-hygiene issues. The manuscript's scope fits cs.CE, but the experimental validation is too thin for the strength of the claims; the missing target-set equivalence is the critical point that must be addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on arXiv:2608.12763. The paper is a straightforward but useful systems integration: a zero-shot VLM front end (DeepSeek-V3 for parsing, Molmo-7B for grounding) feeding a TSP back end that compares the raw VLM order, PSO, and two PSO variants, then picks the shortest route. On the UAV-VLPA-nano-30 benchmark, the full system reports 62.43 km total flight distance, down from 79.66 km for the raw VLA baseline and 69.00 km for human planning, with a 575 s total workflow.\n\nWhat's actually new is the specific combination and the measured result on that benchmark. The individual pieces are known, but the architecture is clean, the paper is readable, and the timing breakdown (VLM dominates at ~19 s/task; TSP ~0.16 s/task) is a useful data point. I credit the authors for seeing that raw VLM ordering is poor and that a cheap TSP post-processor fixes most of it.\n\nThe soft spots are real and they sit in the evaluation. The biggest one: we never see the target sets. No detection precision/recall, no per-task waypoint counts, no statement that ARIES-Mission2, the raw VLA baseline, and the human planner all worked from the same ground-truth targets. For the raw VLA comparison this doesn't matter much, because both methods use the same VLM detections, so the 21.6% reduction is likely a fair routing gain. But the 9.5% margin over human planning is suspect: if the VLM detects fewer or different targets than the human used, the tour shortens for reasons unrelated to routing quality. The HumanPlan protocol is also vague—no interface, instructions, or whether the human saw the same imagery and target definitions.\n\nSecond, the optimization comparisons are single-run. PSO-family variants are stochastic, and the 'optimal' result is the min over four related solvers, which is selection bias if you only report the best. There are no error bars or multiple runs. Third, there is no exact or strong heuristic TSP baseline (e.g., Concorde or LKH). The TSP sizes here are small enough that an exact solver is feasible, and its absence weakens the claim that these metaheuristics are competitive.\n\nThe novelty of GPSO and IPSO individually is minor—linear inertia weight, distance prior initialization, Tent chaos, Levy mutation—but as a systems paper that's acceptable. The citation pattern is heavy on the authors' own PSO variants, which is padding rather than circularity; the benchmark is external and the distances are measured.\n\nNet: this is a reasonable systems paper that would be useful to people building VLA-based drone mission planners. The central performance claim over human planning is not yet falsifiably supported, but the work is not fatally flawed. I'd send it to peer review, not desk-reject, with a required revision that adds detection metrics, multi-run statistics with variances, an exact TSP baseline, and a clear description of the HumanPlan setup. Without those, treat the head-to-head numbers as provisional.","headline":"Useful VLM+TSP systems combo, but the headline gains over human planning are unproven without target-set equivalence and error bars.","tokens_in":9776,"tokens_out":3962,"would_cite":true,"duration_ms":40989,"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":"ARIES-Mission2 shows that separating VLM perception from TSP route optimization generates shorter drone missions than either raw VLA output or human planning, achieving 62.43 km over 30 tasks.","keywords":["vision-language-action","zero-shot target grounding","UAV mission generation","traveling salesperson problem","particle swarm optimization","satellite imagery","MAVLink waypoint generation","metaheuristic route optimization"],"falsifier":"Count the targets the Molmo-7B pipeline detects per task and compare them with the benchmark's ground-truth waypoint set and with the waypoints the human planner visited. If on any task the VLM detects fewer or different waypoints, the reported 9.5% distance gain over human planning is an artifact of comparing different routes, not better planning.","tokens_in":8848,"feed_emoji":"🛸","tokens_out":4967,"duration_ms":46327,"temperature":0.7,"pith_summary":"This paper proposes ARIES-Mission2, a zero-shot framework that generates drone flight missions from natural-language instructions and satellite images by splitting the job in two: a vision-language model finds targets and turns their pixel locations into GPS waypoints, and a separate Traveling Salesperson Problem solver decides the order in which a drone should visit them. The central claim is that this separation fixes a weakness of multimodal models, which can perceive objects but produce wandering, backtracking routes, and that it outperforms both an unoptimized vision-language baseline and human experts. On a 30-task benchmark the full system produces a total route of 62.43 km, compared with 79.66 km for the raw vision-language order and 69.00 km for manual human planning, and it completes all 30 tasks in about 575 seconds. If correct, this indicates that perception and route optimization should be treated as separate modules rather than asked of a single large model.","feed_headline":"One drone-mission pipeline beats human planners by 9.5 percent","feed_subtitle":"A zero-shot model finds targets, a TSP optimizer orders them: 62.43 km vs 69 km human planning across 30 tasks.","key_machinery":"The load-bearing mechanism is the decoupled front-end/back-end architecture with an algorithmic competition selection rule. The front end chains DeepSeek-V3 (instruction parsing) with Molmo-7B (zero-shot target grounding) and maps pixel positions to GPS coordinates by linear interpolation using the image's geographic metadata. The back end formulates the waypoint set as a TSP whose objective is a closed-loop Haversine distance, then runs three particle-swarm variants—standard PSO, GPSO with a distance-based heuristic initialization and linearly decaying inertia weight, and IPSO with Tent-chaos initialization and Lévy-flight mutation—alongside the raw VLM order. The final route is simply the minimum-cost member of these four candidates, encoded as a MAVLink waypoint file locally. The mechanism's contribution is that the route optimization is fully decoupled from perception, so the optimizer's sub-second runtime (0.16 s/task) dwarfs the VLM inference time (19.02 s/task).","core_discovery":"The paper's discovery is a workflow, not a new model: instead of asking a multimodal large language model to both locate targets and order them into a flight path, ARIES-Mission2 lets a language model parse the instruction, a vision model output pixel coordinates, and a deterministic geospatial interpolation convert those coordinates into GPS waypoints. It then treats the remaining multi-target traversal problem as a closed-loop Traveling Salesperson Problem over those waypoints, computes flight costs with the Haversine formula, and competes four candidate visiting orders—the raw VLM order plus routes from PSO, GPSO, and IPSO—selecting the shortest. The result is a 21.6% reduction in total flight distance relative to the pure VLA baseline and a 9.5% reduction relative to an experienced human planner, with the whole 30-task run taking about 575 seconds. The paper presents this as evidence that light metaheuristic optimization, not bigger models, is what closes the gap between semantic understanding and physically efficient flight.","pith_inferences":["A natural next experiment is to report detection precision and recall per task; if recall is high, the distance numbers are strong, and if not, the proper comparison is route cost per detected target.","The same decoupling could be applied to ground robots or maritime vehicles, replacing the Haversine cost with the appropriate metric (road network, currents) and still keeping the perception module unchanged.","One could test whether the four-candidate competition can be replaced by a single well-tuned optimizer or an exact TSP solver; if so, the metaheuristic choice matters less than the separation itself.","The reported scalability analysis suggests a parameter-free prediction: on tasks with more than 20 targets the VLM will dominate runtime even more, so caching or smaller grounding models would improve end-to-end latency more than any solver change."],"forward_implications":["The raw VLM visiting order is the weakest candidate in every reported comparison, so direct VLM route generation is not enough even when perception is accurate.","Because the TSP solver adds only 0.16 s per task, the approach scales to larger missions at negligible optimization cost; the bottleneck is image-to-waypoint grounding.","Generating the final MAVLink mission file locally avoids the latency and hallucination risk of asking a large model to emit control code.","The framework is zero-shot: no fine-tuning or task-specific training is needed to reach the reported route quality.","If the detection parity assumption holds, decoupling perception from combinatorial optimization is a general recipe for spatial-instruction tasks beyond UAVs."],"supporting_citations":[{"why":"Documents the spatial-planning failures of multimodal models in low-altitude UAV scenarios, motivating the need for a separate optimizer.","marker":"[3]"},{"why":"Identifies multi-target UAV traversal as a Traveling Salesperson Problem, the core formulation the back end solves.","marker":"[4]"},{"why":"Supplies the standard particle swarm optimization algorithm that PSO, GPSO, and IPSO are built on.","marker":"[14]"},{"why":"Provides the Lévy-flight mechanism that IPSO uses for nonlinear mutation to avoid premature convergence.","marker":"[15]"},{"why":"Defines the UAV-VLPA-nano-30 benchmark, the dataset of 30 satellite-imagery tasks on which all route-length comparisons are made.","marker":"[17]"}],"fun_headline_variants":["Drone missions: VLM + TSP beats human planning by 9.5%","Zero-shot drone planner: 62 km route, 3.6x faster than humans","VLM for grounding, TSP for routes: 21.6% shorter flights","ARIES-Mission2: zero-shot VLA + TSP for drones"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparisons assume the VLM detects exactly the same set of targets on every task that the human planner used, with no missed targets or hallucinated extra ones; the paper does not report detection recall or waypoint counts per task.","fun_headline_variants_meta":{"raw":{"variants":["Drone missions: VLM + TSP beats human planning by 9.5%","Zero-shot drone planner: 62 km route, 3.6x faster than humans","VLM for grounding, TSP for routes: 21.6% shorter flights","ARIES-Mission2: zero-shot VLA + TSP for drones"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3648,"prompt_tokens":1128,"completion_tokens":2520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":2429}},"tokens_in":744,"tokens_out":2520,"duration_ms":18190,"temperature":1.0,"reasoning_tokens":2429,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:49:30.956836+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count the targets the Molmo-7B pipeline detects per task and compare them with the benchmark's ground-truth waypoint set and with the waypoints the human planner visited. If on any task the VLM detects fewer or different waypoints, the reported 9.5% distance gain over human planning is an artifact of comparing different routes, not better planning.","supporting_citations":[{"cited_title":"Particle swarm optimization[C]//Proceedings of ICNN’95-international conference on neural networks","cited_arxiv_id":null,"evidence_quote":"Supplies the standard particle swarm optimization algorithm that PSO, GPSO, and IPSO are built on."},{"cited_title":"(2025) LSWOA: An enhanced whale optimization algorithm with Levy flight and Spiral flight for numerical and engineering design optimization problems","cited_arxiv_id":null,"evidence_quote":"Provides the Lévy-flight mechanism that IPSO uses for nonlinear mutation to avoid premature convergence."},{"cited_title":"UA V-VLA: Vision-language- action system for large scale aerial mission generation[C]//2025 20th ACM/IEEE International Conference on Human-Robot Interaction (HRI)","cited_arxiv_id":null,"evidence_quote":"Defines the UAV-VLPA-nano-30 benchmark, the dataset of 30 satellite-imagery tasks on which all route-length comparisons are made."}],"review_version":1}