{"id":"896748e4-13d3-4067-84e4-79bc21ab94c3","arxiv_id":"2505.23267","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"VLM-RRT uses a vision-language model to bias RRT sampling toward a suggested direction, reporting fewer iterations and paths close to RRT* in simulated 2D wildfire scenarios.","lead":"The paper adds a vision-language model to the classic RRT path planner: the model looks at a picture of the map and says which of eight directions to search next. The reported gain is fewer RRT iterations to find a path, but the method is only tested in simulation and the extra time cost of the model is not measured.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"VLM-RRT's gains may come from the sector sampler itself, not the VLM; the missing VLM-ablation is the load-bearing gap.","rationale":"The reader's weakest assumption targets the VLM's directional accuracy and its reliability at the eight-direction granularity. My concern is adjacent but more specific: the experiments do not isolate whether the VLM output is the cause of the reported improvement. The paper compares VLM-RRT against RRT and RRT*, but never against a version of VLM-RRT that uses the same sector-sampling machinery without a VLM, or with a trivial goal-direction oracle. Because the goal is explicitly marked in the image and the VLM is asked to choose among only eight directions, a simple direction-to-goal heuristic may reproduce most of the benefit—making the VLM's 'pattern matching' unnecessary. The γ = 1 result in Table III strengthens this worry: when the VLM is always consulted, success degrades below plain RRT, so the VLM is not a reliable guiding signal. The missing ablation is therefore the single most load-bearing gap: it determines whether the method is a genuine VLM-based planner or a sector-biased RRT with an expensive direction source. I agree with the reader that VLM reliability is unverified, but I would frame the required check as an ablation that separates the sampling mechanism from the VLM. The reader's conditional verdict remains appropriate: the paper should be accepted only if this ablation, along with the already-noted statistical reporting and code release, is provided. I do not see a reason to move to a stronger or weaker verdict based on this specific concern, so I recommend UNCHANGED.","tokens_in":12398,"tokens_out":3868,"duration_ms":47828,"concrete_test":"Run the same 250-scenario Monte Carlo evaluation used for Table II with the VLM replaced by two control policies, keeping all other parameters fixed (γ = 0.85, r = 30 m, θ = 45°, δ, ϵ, N = 500): (a) a uniformly random direction sampled from the eight allowed labels, and (b) the direction from the selected leaf node ν̂ toward the goal centroid Go (computed directly from the environment state). Compare success rate, average iterations, and average path length against VLM-RRT with GPT-4o few-shot and Llama 3.2V few-shot. As an additional check, measure the VLM's per-output accuracy against a ground-truth optimal direction computed from the same rendered environment. If control (b) matches or beats VLM-RRT's success rate and iteration count, the VLM's reasoning contributes nothing beyond a goal-biased sampler; if control (a) also matches, sector sampling alone explains the gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—that VLM guidance improves sampling efficiency—rests on attributing the iteration reductions in Table II to the VLM's directional output (Algorithm 2, Lines 6–8). But the experimental design never ablates the VLM: the sector-based sampler (radius r = 30 m, angle θ = 45°) is always paired with a VLM output, so the observed gains could come from the local sector sampling mechanism itself, not from the VLM's pattern recognition or reasoning. This is not merely hypothetical: Table III shows that at γ = 1 (VLM consulted at every step) the success rate drops to 79%, below plain RRT's 82%, and the text attributes this to 'the VLM can make mistakes.' Thus the VLM's directional advice is not reliably better than the uniform sampling it replaces. Without a control condition that replaces the VLM with (a) a uniformly random direction among the eight choices and (b) a trivial heuristic that points from the selected leaf toward the goal centroid (which is explicitly marked as a red dot in the rendered image), the reported superiority of VLM-RRT over RRT cannot be attributed to the VLM. Since the paper's stated novelty is exactly this VLM integration, this missing ablation is load-bearing.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes VLM-RRT, a hybrid path-planning method that integrates a vision-language model (VLM) into the RRT sampling loop. At each iteration, with probability γ, the VLM is shown a rendered environment image (including start, goal, obstacles, and the current tree) and is asked to output one of eight cardinal/intercardinal directions; the planner then samples a new point from a sector centered at a randomly selected leaf node, with radius r and angle θ oriented along the VLM's suggested direction. The algorithm is evaluated in a 2D UAV navigation setting, comparing GPT-4o and Llama 3.2V under zero-shot, few-shot, and chain-of-thought prompting against RRT, RRT*, and LLM-A*, reporting success rate, iteration count, and path length, plus a robustness sweep over γ and a dynamic-goal experiment.","tokens_in":12609,"tokens_out":3636,"duration_ms":38423,"significance":"If the claimed gains are real, VLM-guided sampling would be a novel and practically relevant way to inject semantic perception into sampling-based planners, and the paper's systematic comparison across two VLMs and three prompting schemes is a useful starting point. However, the evidence as presented is not yet load-bearing: the central attribution of the improvements to the VLM is untested because the sector sampler is never ablated, the quantitative claims are weakened by inconsistent numbers across tables and the absence of variance or significance measures, and the practical-efficiency motivation is not supported because wall-clock time and VLM API latency are never reported. The γ robustness sweep (Table III) is a commendable attempt to characterize sensitivity, but it also exposes that the VLM's direction advice is sometimes harmful, reinforcing the need for the missing control conditions.","major_comments":[{"comment":"The central claim that VLM guidance improves sampling efficiency is not supported because no control condition ablates the VLM. The sector sampler (radius r=30 m, angle θ=45°) is always paired with the VLM's directional output, so the observed iteration reductions in Table II could be produced by the local sector sampling mechanism alone, irrespective of VLM accuracy. This concern is sharpened by Table III: at γ=1, where the VLM is consulted at every step, success rate drops to 79%, below plain RRT's 82%, and the text attributes this to 'the VLM can make mistakes' (Section V-B). To attribute the gains to the VLM, the authors should add control conditions that replace the VLM direction with (a) a uniformly random choice among the eight directions and (b) a trivial goal-pointing heuristic (e.g., the direction from the selected leaf toward the goal centroid, which is explicitly marked as a red dot in the rendered image). Without these conditions, the paper's novelty claim is not empirically established.","section":"Section V-B and Algorithm 2 (Lines 6-8)"},{"comment":"The paper reports inconsistent averages for the same algorithm and prompting configuration. For GPT-4o with CoT prompting, Table I reports an average of 172 iterations and 54.56 m path length, while Table II reports 94 iterations and 48 m for the identical configuration; similar discrepancies appear for RRT (423/56.48 vs 343/58) and RRT* (477/53.89 vs 302/45). The text states that Table I averages 100 MC runs and Table II averages 250 MC runs, but a difference of this magnitude indicates different environment distributions, termination criteria, or success-rate filtering, and the authors do not explain it. Additionally, the paper reports no variance, confidence intervals, or significance tests, so the claim of 'significantly fewer iterations' (Section V-B) is not statistically supported.","section":"Tables I and II"},{"comment":"No wall-clock time or end-to-end latency is reported. The paper motivates the work by 'rapid and efficient' navigation, but each VLM-informed step in Algorithm 2 requires a rendered image and an API call to GPT-4o or Llama 3.2V; the iteration-count reductions in Tables I-III may be offset by VLM inference latency. At minimum, the authors should report planning wall-clock time with and without the VLM, including API latency, and ideally the number of VLM calls per successful plan. Without this, the practical-efficiency claim is incomplete.","section":"Section V-A and V-B"}],"minor_comments":[{"comment":"The final constraint 'y(t) ∈ Y, ∀k ∈ {0,...,N−1}' uses the index k instead of t, and the horizon upper limit should be consistent with T−1 rather than N−1.","section":"Equation (2)"},{"comment":"The phrase 'The Trajectory also serve as [Previous state]' has a subject-verb agreement error and missing words; please revise.","section":"Fig. 1 caption"},{"comment":"The statement that VLM-RRT 'significantly' improves performance should be supported by statistical tests or at least variance measures; the zero-shot conditions (GPT-4o success rate 68% vs RRT's 82%) show that the VLM can degrade performance, so the claimed improvement is conditional on prompting strategy.","section":"Section V-B"},{"comment":"The notation r is used both for the sector radius in SampleStateVLM and for the empty set initialized as 'r ← ∅' in Line 1; rename one of these to avoid ambiguity.","section":"Algorithm 2, Line 1 and Line 8"},{"comment":"The reference path P is used in the cost function of Eq. (2) before it is formally defined as the output of the planner; consider defining P in the problem formulation or reordering the presentation so that the planning stage is described first.","section":"Section III-C"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between Table I and Table II for identical algorithm/prompt configurations is a substantive concern that goes beyond ordinary Monte Carlo noise and should be explicitly reconciled by the authors. The missing VLM ablation is the key technical gap that currently prevents the paper from supporting its novelty claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper does something modest and mostly sensible—use a VLM to pick one of eight directions from a rendered top-down image, then bias RRT sampling into a sector around that direction. This is a direct extension of learned-guidance RRT (Neural RRT*, LLM-A*), and the paper is honest about that lineage. The prompting study (zero-shot, few-shot, CoT; GPT-4o vs Llama 3.2V) is useful, and the robustness sweep on gamma is a nice touch.\n\nWhat I like: the algorithm is simple and clearly specified. The evaluation covers success rate, iterations, path length, dynamic goals, and gamma sensitivity. The authors openly show that gamma=1 hurts (Table III, success drops to 79% vs 82% for plain RRT) and explain it as VLM mistakes plus greediness. That is good faith.\n\nThe soft spots are real, though. The load-bearing gap is the missing VLM ablation. The sector sampler (r=30 m, theta=45°) is always paired with VLM direction; there is no control that replaces the VLM with (a) a random one of the eight directions or (b) a trivial \"point toward goal centroid\" heuristic. Because the VLM's advice at gamma=1 is sometimes worse than uniform sampling, the iteration gains in Table II may come from the sector mechanism itself, not the VLM. That is exactly the paper's claimed novelty, so it needs fixing.\n\nAlso, Table I and Table II report different numbers for the same condition (GPT-4o CoT: 172 iters / 54.56 m vs 94 iters / 48 m). Possibly different MC trial counts, but the paper doesn't say, and that erodes trust. There are no error bars or significance tests; the phrase \"significantly fewer iterations\" is doing work without statistics. No wall-clock time or VLM API latency, which matters for a real-time UAV claim. Zero-shot success rates (56–68%) are worse than plain RRT (82%)—worth stating more prominently. No code or data.\n\nThe math and citation pattern look fine. The method is empirical and the paper doesn't overclaim a derivation. Related work is appropriate; self-citations are not the issue.\n\nBottom line: this is a plausible, incremental contribution that a workshop or short-conference audience would learn from. A serious referee should see it, because the missing ablation is fixable and the problem setup is clean. My recommendation: send it to review, but the authors must add the VLM-free control conditions, report variance and runtime, and reconcile the tables.","headline":"VLM-RRT is a plausible incremental idea whose main claim needs a VLM-free control before the iteration gains can be credited to the VLM.","tokens_in":13183,"tokens_out":2139,"would_cite":false,"duration_ms":22137,"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":"This paper claims that a vision-language model can serve as a directional prior inside an RRT path planner, cutting the number of sampling iterations to about a third while preserving path quality.","keywords":["VLM-RRT","UAV navigation","sampling-based path planning","vision language models","Rapidly-exploring Random Trees","chain-of-thought prompting","wildfire disaster response"],"falsifier":"Run VLM-RRT on the same benchmark with the VLM's direction replaced by a random draw from the same eight compass directions, and also by the straight-line direction to the goal, keeping the sector radius and angle fixed; if success rate and iteration counts do not clearly worsen, the claimed gains come from the sector sampling geometry rather than the model's visual reasoning.","tokens_in":12156,"feed_emoji":"🚁","tokens_out":9147,"duration_ms":96689,"temperature":0.7,"pith_summary":"The paper sets out to show that a vision-language model can make sampling-based path planning much cheaper by telling the sampler where to look next. In the proposed VLM-RRT planner, the VLM is shown a rendered image of the environment with the start, goal, obstacles, and the partially built tree marked, and it answers with a single compass direction. The RRT loop then draws its next random sample from a small sector in that direction most of the time, instead of uniformly across the whole map. The authors report that this guided sampling finds collision-free paths in far fewer iterations than plain RRT and RRT*, with success rates above both when the prompt includes worked examples or chain-of-thought reasoning, and with path lengths close to RRT*'s. If this holds up, the practical payoff is a simple way to inject visual common sense into planners used for time-critical UAV missions such as wildfire search and rescue.","feed_headline":"Vision-model hint makes RRT planning three times faster","feed_subtitle":"Guided sampling beats uniform random search on success rate and path length in simulated UAV missions.","key_machinery":"The load-bearing mechanism is VLM-guided sector sampling: instead of drawing the next sample uniformly from the whole environment, the planner selects a leaf node, asks the vision-language model for one of eight directions toward the goal, and draws a sample uniformly from a sector $R$ of radius 30 m and angle 45 degrees centred at that leaf and aligned with the suggested direction. The parameter $\\gamma$, defaulting to 0.85, decides how often this guided draw replaces uniform sampling, so the algorithm keeps an exploration fallback. The prompt system, with zero-shot, few-shot, and chain-of-thought variants, is what elicits the directional output from the VLM, and the evaluation shows the few-shot and chain-of-thought variants are the ones that make the planner outperform the baselines.","core_discovery":"The central claim is that VLM-RRT beats traditional RRT by treating a vision-language model as a general-purpose pattern matcher that supplies a coarse directional heuristic. At each guided step the algorithm picks a leaf of the growing tree, renders the current environment with the goal, obstacles, and explored tree marked, asks the VLM for one word from eight compass directions, and then samples a new point inside a 45-degree sector of radius 30 metres centred on that leaf and oriented along the suggested direction. With a consultation probability of $\\gamma=0.85$, the planner alternates between this guided sampling and ordinary uniform RRT sampling. The paper's experiments across 250 random wildfire-inspired scenarios report that this hybrid finds paths in about one-third of the iterations needed by RRT or RRT*, with success rates up to 94% versus 82% for RRT and 88% for RRT*, and with path length essentially matching RRT*.","pith_inferences":["An ablation that replaces the VLM's answer with a fixed rule such as 'sample in a cone toward the goal' would reveal how much of the speed-up is due to the visual reasoning itself rather than to the sector-biased sampling scheme.","The eight-word direction vocabulary makes VLM-RRT a discrete local search; coupling it with rewiring, as in RRT*, could turn the iteration savings into asymptotically optimal paths, which the paper does not attempt.","The strong dependence of success rate on prompt style suggests that prompt engineering matters as much as the planner's own parameters, and that a deployable system would need a way to detect and recover from VLM misdirections rather than trusting a fixed $\\gamma$."],"forward_implications":["With the best prompt style, VLM-RRT finds collision-free paths in roughly 89 iterations on average, compared with 343 for RRT and 302 for RRT*, on the same 250-scenario benchmark.","The improvement does not cost path quality: the best VLM-RRT average path length is 46 m, close to RRT*'s 45 m and well under RRT's 58 m.","The guided planner can react to a goal that moves during the mission, detecting the relocation in 92% of 50 dynamic scenarios and usually correcting its direction within one extra sampling step.","Because $\\gamma$ can be reduced, the planner can degrade smoothly toward ordinary RRT when the VLM's advice is not trusted, trading speed for robustness."],"supporting_citations":[{"why":"Defines the baseline RRT algorithm that VLM-RRT augments and that the experiments compare against.","marker":"[23]"},{"why":"Provides the asymptotically optimal RRT* variant used as a reference for iteration count and path length.","marker":"[28]"},{"why":"Supplies the LLM-A* approach, the closest prior method that also fuses a language model with a classical planner.","marker":"[38]"},{"why":"Supplies the A* grid-search baseline in the first performance comparison table.","marker":"[39]"},{"why":"Introduces chain-of-thought prompting, one of the three prompt techniques whose results are compared.","marker":"[42]"},{"why":"Reports the proprietary multimodal model used as the VLM module in the main experiments.","marker":"[44]"},{"why":"Reports the open-weights vision model used as the alternative VLM module to test whether the result transfers across models.","marker":"[45]"}],"fun_headline_variants":["VLM-guided RRT cuts planning iterations by two-thirds","Vision language model guides RRT to faster, safer UAV paths","Hybrid planner: VLM steering speeds RRT threefold","VLM gives RRT a directional nudge for quicker paths","AI vision model boosts RRT success to 94% in UAV tests"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The result rests on the assumption that the VLM's one-word direction is reliable enough at the eight-direction level to bias sampling usefully; the paper's own robustness test shows that when every step follows the VLM ($\\gamma=1$), its mistakes lower success to 79%, below plain RRT's 82%.","fun_headline_variants_meta":{"raw":{"variants":["VLM-guided RRT cuts planning iterations by two-thirds","Vision language model guides RRT to faster, safer UAV paths","Hybrid planner: VLM steering speeds RRT threefold","VLM gives RRT a directional nudge for quicker paths","AI vision model boosts RRT success to 94% in UAV tests"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000268,"raw_usage":{"total_tokens":1601,"prompt_tokens":911,"completion_tokens":690,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":603}},"tokens_in":527,"tokens_out":690,"duration_ms":7114,"temperature":1.0,"reasoning_tokens":603,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:48:33.446272+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run VLM-RRT on the same benchmark with the VLM's direction replaced by a random draw from the same eight compass directions, and also by the straight-line direction to the goal, keeping the sector radius and angle fixed; if success rate and iteration counts do not clearly worsen, the claimed gains come from the sector sampling geometry rather than the model's visual reasoning.","supporting_citations":[{"cited_title":"Rapidly-exploring random trees: A new tool for path planning,","cited_arxiv_id":null,"evidence_quote":"Defines the baseline RRT algorithm that VLM-RRT augments and that the experiments compare against."}],"review_version":1}