{"id":"085b0762-8d1a-4545-98e5-1e46e0ce6eb5","arxiv_id":"2508.20884","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A new sampling-based planner, LIT*, dynamically tunes batch size and nearest-neighbor count using a fuzzy DDPG tensor, and claims faster convergence and lower path cost in R4-R16 and dual-arm manipulation.","lead":"LIT* is a robot motion planner that uses a fuzzy reinforcement learning model to adjust how many samples and neighbors it considers based on obstacle density. The paper reports faster paths and lower costs in high-dimensional tests, which could make robot arms plan better in cluttered spaces.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Fuzzy-DDPG training distribution is unspecified; if the NP/RR test maps were used during training, the reported gains reduce to memorization, not adaptation.","rationale":"The reader's weakest_assumption identifies the same load-bearing concern I judge most critical: the learned tensor's generalization is unverified because the training distribution is unspecified. Algorithm 3's resetEnvironment() and Section III.B provide no training/test separation, and Section IV evaluates on NP/RR maps without stating they were unseen. Since the tensor is fixed at deployment, the claimed adaptability rests entirely on offline training transfer. I also noticed that Table I's 31.59% value corresponds to the tmed-init improvement column, not cost improvement, so the Section IV sentence describing it as a cost reduction is inaccurate; this further weakens the summary but does not change the fundamental conditional recommendation. The proposed held-out cross-distribution test would directly settle whether the central claim of adaptation survives.","tokens_in":10354,"tokens_out":8129,"duration_ms":85433,"concrete_test":"Perform a held-out generalization test: train Fuzzy-DDPG on one distribution (e.g., Random Rectangles maps only) and evaluate on a disjoint distribution (e.g., Narrow Passage maps only), and the reverse, with the same 100-run protocol as Section IV. Report training environments, episode count, and hyperparameters. If the tensor trained on one map distribution fails to match or beat FIT* on the other, the reported gains are explained by training/evaluation overlap rather than adaptive B/K selection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the Fuzzy-DDPG tensor learned during training generalizes to the Section IV benchmarks. This premise is unsecured. Algorithm 3 and Section III.B specify only a generic DDPG loop (resetEnvironment, replay buffer, soft updates); the paper never states which maps were used for training, how many episodes, what hyperparameters (γ, τ, learning rates, reward scales), or whether the NP, RR, and dual-arm maps were held out. The deployed artifact is a fixed tensor (Fig. 5) queried in O(1), so all adaptation is frozen before evaluation. If the training environments included the same NP/RR scenarios reported in Table I, the improvements could be an artifact of memorizing those obstacle layouts. No train/test split or held-out environment is described, so the paper's 'environmental adaptability' claim is not empirically supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LIT*, a sampling-based motion planner that learns to adapt two parameters of informed-sampling tree planners: the batch size B and the neighbor-count factor ψK (with K computed from Eq. 4). Environmental context is encoded by global invalid ratio, local invalid ratio, and Lebesgue measure of the informed set, fuzzified into a 9-dimensional representation and fed to a DDPG-based actor-critic whose defuzzified outputs are pre-mapped into a lookup tensor. During deployment, B and ψK are retrieved in O(1) from this fixed tensor. The authors evaluate LIT* against RRT-Connect, Informed RRT*, AIT*, BIT*, EIT*, and FIT* in Narrow Passage and Random Rectangles benchmarks from R4 to R16, plus a dual-arm robot manipulation task, reporting up to 31.59% lower median initial cost relative to FIT*.","tokens_in":10740,"tokens_out":2625,"duration_ms":31213,"significance":"If the empirical claims hold, the paper would make a useful contribution: it couples a fuzzy encoding of obstacle distribution with reinforcement learning to adapt batch size and neighbor count, and the tensor-lookup deployment is computationally attractive for repeated planning. The idea of using invalid sampled states to modulate both exploration density and local connectivity is intuitive and goes beyond fixed-parameter informed planners. However, the current manuscript does not provide enough information about the training procedure, environments, or hyperparameters to establish that the learned tensor generalizes. The paper also lacks quantitative comparisons against most claimed baselines. These omissions are load-bearing because the central claim is environmental adaptability, not merely the introduction of a new planner variant.","major_comments":[{"comment":"The training setup is underspecified. Algorithm 3 calls resetEnvironment() and loops over episodes, but the manuscript never states which maps are used for training, how many episodes are run, what DDPG hyperparameters are used (γ, τ, minibatch size, replay buffer size, PER exponent), or whether the NP/RR/dual-arm maps in Section IV were held out. Since deployment uses a fixed, precomputed tensor (Fig. 5) queried in O(1), all adaptation is frozen before evaluation. If the same maps were used during training, the reported gains in Table I could reflect memorization rather than generalization. The authors must report a train/test split, include held-out environments, and provide all fuzzification parameters (ϑ, σ), reward scales, and tensor discretization details.","section":"Section III.B and Algorithm 3"},{"comment":"The abstract and conclusion claim LIT* 'outperforms state-of-the-art single-query, sampling-based planners,' but Table I quantifies improvements only against FIT*, the authors' own prior method. No numerical success rates, cost statistics, or statistical tests are reported for RRT-Connect, Informed RRT*, AIT*, BIT*, or EIT*. The plots in Fig. 6 show median curves with confidence intervals, but the reader cannot determine, for example, whether LIT* is statistically better than BIT* or EIT* in each scenario. A complete comparison table with success rates, medians, and confidence intervals for all planners is needed to support the central claim.","section":"Section IV, Table I and Fig. 6"},{"comment":"The reward functions directly minimize the same metrics used in evaluation (time and cost), which is not itself a flaw, but the scaling factors αB, βB, γB, αK, βK, γK are never reported, and the damping function κ(nupdate) = max(νmin, ν·log2(6.8 − nupdate)) is not fully specified: its domain and behavior for nupdate ≥ 6.8 are undefined, and ν, νmin are not given. Without these values, the learned policy cannot be reproduced, and it is impossible to tell whether the reported behavior comes from the learning algorithm or from hand-tuned reward shaping.","section":"Section III.B.1, Eqs. (9)-(10)"}],"minor_comments":[{"comment":"Eq. (13) defines δ = Qφ′(s,a) − Qφ(s,a), which is not the TD error used in Eq. (8); the target should involve the reward and target networks. Eq. (15) updates W′_φ using W_θ, which appears to be a typo for W_θ′. Please correct these equations and the surrounding notation.","section":"Section III.A.2, Eqs. (13)-(16)"},{"comment":"There are several typos and inconsistent notations: 'Specificlly', 'Leaning-based', 'Lebesgue Messure', 'caucLocalRadius', 'Adoption' in Fig. 1, and 'T rain' in the abstract. In Table I the percentage column headers are ambiguous; please label whether they refer to t_med_init, c_med_init, or c_med_final. Also, the relationship between the 9×1 fuzzified vector and the convolutional layer with kernels 3, 5, and 7 could be explained more clearly.","section":"Throughout"},{"comment":"The discussion mentions communication latency between C++ and Python as a limitation but does not quantify its impact on the reported computation times. A sentence clarifying whether evaluation time includes tensor lookup and inter-process communication would help interpret the runtime results.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The manuscript relies heavily on self-citations (references [11], [12], [17], [18], [19], [20]), and the novelty relative to FIT* and APT* should be scrutinized by the editor. The absence of code release and training details may also be a reproducibility concern for the journal. The empirical claim of environmental adaptability is currently unsupported without a train/test split."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: LIT* is a real algorithmic variant, but the paper's central empirical claim is not currently supported because the training distribution is undisclosed and the evaluation reuses the same benchmark family. The stress-test note is correct.\n\nWhat's new: LIT* combines batch-size and neighbor-count adaptation in one fuzzy-DDPG controller, using global/local invalid ratios and Lebesgue measure as features. That combination is not in the prior literature. The O(1) tensor lookup at runtime is a sensible way to deploy a learned policy inside a C++ planner. Experiments span R4-R16 in narrow passage and random rectangles, plus a dual-arm OpenRAVE task, and Table I shows consistent double-digit percentage reductions in median initial cost over FIT*, their own prior method.\n\nThe soft spots are substantial. The paper never states which environments were used to train the DDPG, nor the fuzzification parameters (ϑ, σ), reward scaling factors, tensor discretization, or DDPG hyperparameters. No code is released. Table I only quantifies gains against FIT*; comparisons against AIT*, BIT*, and EIT* appear in plots but not in the table. More importantly, the reward functions (Eqs. 9 and 10) directly minimize time and cost, the same metrics used in evaluation, and the trained tensor is frozen at test time. Without a held-out environment or explicit train/test split, the reported 31.59% improvement on NP-R16 could simply reflect memorizing those obstacle layouts. That is the load-bearing flaw, and it is fixable: release training details and evaluate on maps not seen during training.\n\nThe method itself is coherent. The writing is rough in places, and the C++/Python communication overhead is acknowledged but not measured.\n\nBottom line: read this if you work on learned parameter adaptation in sampling-based planning. It deserves a serious referee, but only with major revisions that include held-out evaluation and full hyperparameter disclosure. I would not cite it in its current form.","headline":"LIT* is a plausible fuzzy-RL parameter scheduler for BIT*-style planners, but missing training details and an absent train/test split make the headline gains over FIT* uninterpretable.","tokens_in":11117,"tokens_out":3375,"would_cite":false,"duration_ms":33648,"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 learning-based planner, LIT*, that dynamically adjusts batch size and nearest-neighbor count via a fuzzy reinforcement learning policy, finds lower-cost paths and converges faster than current single-query sampling-","keywords":["sampling-based motion planning","deep fuzzy reinforcement learning","adaptive batch size","k-nearest neighbor selection","informed trees","high-dimensional planning","invalid-state ratio","dual-arm manipulation"],"falsifier":"Take the frozen B and K tensors and run LIT* on a set of narrow-passage and random-rectangle maps that were not used during training, with all baselines given the same time limits; if the median initial-cost advantage over FIT* shrinks to noise, the central claim of learned adaptability fails. A sharper version would train only on R4 maps and evaluate on R8 and R16 maps.","tokens_in":10332,"feed_emoji":"🤖","tokens_out":7443,"duration_ms":73633,"temperature":0.7,"pith_summary":"LIT* is a sampling-based motion planner that treats two internal parameters—the number of samples per batch and the number of neighbors used to extend the tree—as learned controls rather than fixed constants. The planner fuzzifies three environmental signals (global invalid-state ratio, local invalid-state ratio around the current solution path, and the volume of the informed ellipsoid) and feeds them to a DDPG-trained actor that outputs fuzzy membership values; defuzzification yields concrete batch size and neighbor-factor values. The central claim is that this parameter adaptation to obstacle density improves both convergence speed and initial solution quality, with reported median initial-cost reductions of up to 31.59% on a narrow-passage R16 benchmark and a successful demonstration on a 14-DOF dual-arm manipulation task. The paper's insight is that invalid sampled states are a cheap, informative measure of environmental difficulty, and that spending more samples or more neighbors exactly where they are needed is better than any fixed trade-off.","feed_headline":"Learned fuzzy tuning cuts path cost by 31.6%","feed_subtitle":"A sampling planner adapts batch size and neighbor count to obstacle density, pulling ahead of fixed-parameter rivals in R8–R16.","key_machinery":"The carrying mechanism is the Fuzzy-DDPG parameter tensor. Three features—the global invalid ratio ρ = |X_invalid|/(|X_valid| + |X_invalid|), the local invalid ratio computed in a radius around the current path and expansion center, and the Lebesgue measure of the informed set—are fuzzified with Gaussian membership functions into a 9-dimensional vector. An actor-critic DDPG network, with TSK defuzzification, maps this vector to crisp values B in [20, 200] and ψK in [3.0, 15.0]; K is then computed from ψK by Eq. 4. After training, the whole mapping is precomputed into a three-axis tensor, so inference at plan time is a single table lookup. The invalid ratio is the key information carrier: it","core_discovery":"The paper's central discovery claim is that adaptive, learned parameter setting outperforms the hand-designed or fixed parameter settings of prior informed-tree planners. Concretely, LIT* extends FIT*'s adaptive batch-size idea by adding a learned neighbor factor ψK, so that K = η e ψK (1 + 1/n) log |X_valid|. Both batch size B and ψK come from a fuzzy-DDPG policy whose observation is a 9-dimensional fuzzified vector built from the global invalid ratio, the local invalid ratio, and the informed-set measure. After training, the policy is frozen into a 3D lookup tensor indexed by the three features, so runtime queries are O(1). Evaluated over 100 runs on narrow-passage and random-rectangle map","pith_inferences":["The paper does not describe the Fuzzy-DDPG training environments or any train/test split; the most direct test of the contribution is whether the frozen tensor retains its advantage on maps never seen during training.","Since the tensor encodes only three scalar features, the learned mapping might generalize across problem classes with similar obstacle statistics; that transfer claim is not made by the paper but is testable.","The same fuzzified invalid-ratio encoding could plausibly control other resource parameters of sampling planners, such as rewiring radius or sampling-proposal variance, which the paper leaves implicit.","The acknowledged C++/Python shared-file communication is a training-time bottleneck; replacing it with direct in-process calls would strengthen the reported wall-clock gains."],"forward_implications":["High-dimensional planning (R8–R16) is where the reported gains are largest, suggesting that fixed batch and neighbor settings become a real bottleneck as the configuration-space dimension grows.","The O(1) tensor lookup means the learned parameter policy adds only microsecond-level overhead at runtime, preserving the anytime property of the planner.","Because the three input features are cheap to compute for any informed sampling-based planner, the learned batch-size and neighbor selection could be transferred to other batch-based planners.","The successful R14 dual-arm manipulation experiment indicates that the mechanism is not limited to synthetic random maps and may hold for real manipulation tasks.","Reward shaping that rewards fast initial solutions and short paths should make LIT* suitable for online replanning and time-critical manipulation."],"supporting_citations":[{"why":"Supplies the batch-informed tree structure that LIT* builds on and uses as a baseline.","marker":"[9]"},{"why":"Direct predecessor and main baseline; contributes the adaptive batch-size mechanism that LIT* extends with learned neighbor selection.","marker":"[19]"},{"why":"Provides AIT* and EIT* baselines and the heuristic-informed region machinery used by the planner family.","marker":"[10]"},{"why":"Recent adaptive neighbor-selection baseline whose fixed-elipse approach motivates the learned K factor.","marker":"[18]"},{"why":"Introduces the admissible informed ellipsoid whose volume is used as an input feature to the fuzzy policy.","marker":"[7]"},{"why":"Establishes the asymptotic optimality framework that the informed-tree planners inherit.","marker":"[6]"},{"why":"Supplies the benchmark harness and confidence-interval statistics used in all experimental comparisons.","marker":"[24]"},{"why":"Provides the baseline planner implementations used in the comparison.","marker":"[29]"},{"why":"Prioritized experience replay is used in training to handle sparse reward signals.","marker":"[28]"}],"fun_headline_variants":["LIT*: Fuzzy deep learning adapts batch and neighbor count for speed","Self-tuning planner uses fuzzy policy to cut path cost in high-D","Adaptive fuzzy planner beats fixed-parameter rivals in R8-R14","Deep fuzzy planner cuts compute time by adapting to obstacle density"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the Fuzzy-DDPG policy, whose training maps and hyperparameters are not reported, learns to generalize rather than memorizes the benchmark maps; without a shown train/test split, the reported improvements could come from tuning to the evaluation environments.","fun_headline_variants_meta":{"raw":{"variants":["LIT*: Fuzzy deep learning adapts batch and neighbor count for speed","Self-tuning planner uses fuzzy policy to cut path cost in high-D","Adaptive fuzzy planner beats fixed-parameter rivals in R8-R14","Deep fuzzy planner cuts compute time by adapting to obstacle density"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001548,"raw_usage":{"total_tokens":6024,"prompt_tokens":740,"completion_tokens":5284,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":5208}},"tokens_in":484,"tokens_out":5284,"duration_ms":37527,"temperature":1.0,"reasoning_tokens":5208,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:43:43.421375+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the frozen B and K tensors and run LIT* on a set of narrow-passage and random-rectangle maps that were not used during training, with all baselines given the same time limits; if the median initial-cost advantage over FIT* shrinks to noise, the central claim of learned adaptability fails. A sharper version would train only on R4 maps and evaluate on R8 and R16 maps.","supporting_citations":[{"cited_title":"Batch informed trees (BIT*): Informed asymptotically optimal anytime search,","cited_arxiv_id":null,"evidence_quote":"Supplies the batch-informed tree structure that LIT* builds on and uses as a baseline."},{"cited_title":"Flexible informed trees (FIT*): Adaptive batch-size approach in informed sampling- based path planning,","cited_arxiv_id":null,"evidence_quote":"Direct predecessor and main baseline; contributes the adaptive batch-size mechanism that LIT* extends with learned neighbor selection."},{"cited_title":"Adaptively informed trees (AIT*) and effort informed trees (EIT*): Asymmetric bidirectional sampling- based path planning,","cited_arxiv_id":null,"evidence_quote":"Provides AIT* and EIT* baselines and the heuristic-informed region machinery used by the planner family."},{"cited_title":"Informed RRT*: Optimal sampling-based path planning focused via direct sampling of an admissible ellipsoidal heuristic,","cited_arxiv_id":null,"evidence_quote":"Introduces the admissible informed ellipsoid whose volume is used as an input feature to the fuzzy policy."},{"cited_title":"Planner developer tools (pdt): Reproducible experiments and statistical analysis for developing and testing motion planners,","cited_arxiv_id":null,"evidence_quote":"Supplies the benchmark harness and confidence-interval statistics used in all experimental comparisons."}],"review_version":1}