{"id":"83efd958-7d84-4053-8639-5e83cfdb7458","arxiv_id":"2607.12370","paper_version":1,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Dual-stream Mamba with fast-decay reactive and slow-decay strategic memory improves path efficiency and timeout rates for LiDAR-based robot navigation over single-stream temporal baselines.","lead":"StratMamba is a dual-stream Mamba model that splits LiDAR memory into fast reactive and slow strategic streams for robot obstacle avoidance. It reports better path efficiency and timeout rates than LSTM, Transformer, and vanilla Mamba in sim and on a Unitree GO1.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.5","headline":"Abstract-only review leaves the dual-stream causal claim untestable; no stronger load-bearing flaw is identifiable beyond the reader's matched-capacity concern.","rationale":"The Reader correctly flags the unmatched-capacity / missing-statistical-isolation issue as the weakest assumption and assigns CONDITIONAL with LOW confidence on abstract-only evidence. Full text is unavailable, so no additional technical soft spot (hidden assumption in the decay formulation, metric definition error, sim-to-real protocol flaw, etc.) can be located. The metrics themselves are concrete and the dual-stream motivation is standard multi-timescale reasoning; disagreement with consensus is not at issue. Therefore the verdict stays CONDITIONAL and agreement is full. The concrete test simply operationalizes the Reader's concern once the paper is released.","tokens_in":2065,"tokens_out":450,"duration_ms":4037,"concrete_test":"When the full paper appears, extract parameter counts and training budgets for StratMamba vs Vanilla-Mamba; re-train or re-evaluate a capacity-matched Vanilla-Mamba (or single-stream ablation) under the same seed set and report mean±std of path efficiency and median steps. If the dual-stream advantage shrinks below ~2% or loses significance, the architectural claim weakens.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant objection beyond the reader's weakest assumption can be substantiated from the abstract alone. The central claim (dual-stream fast/slow decay Mamba yields 0.915 path efficiency, 576 median steps, lower timeouts vs LSTM/Transformer/Vanilla-Mamba, plus GO1 sim-to-real) is internally consistent and empirically framed. The load-bearing condition is that the reported gains are produced by the strategic/reactive stream partition rather than unmatched capacity, training budget, or variance. The abstract supplies no parameter counts, FLOPs, identical hyperparameter schedules, ablation removing only the dual-stream split, or error bars/statistical tests. Without the full paper this condition cannot be verified, but nothing in the abstract contradicts it or introduces a deeper inconsistency (e.g., circular metrics or impossible numbers).","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The manuscript proposes StratMamba, a dual-stream Mamba temporal architecture for LiDAR-based obstacle avoidance that partitions memory into a fast-decay reactive stream (high-frequency LiDAR for local avoidance) and a slow-decay strategic stream (longer-horizon goal information). Evaluations in IsaacLab and Gazebo, plus sim-to-real deployment on a Unitree GO1 with static and dynamic obstacles, are claimed to show lower timeout rates, fastest navigation (576 median steps, 5.0% better than Vanilla-Mamba), and highest path optimality (0.915 path efficiency) relative to LSTM, Transformer, and Vanilla-Mamba baselines, with more robust performance under extended LiDAR ranges.","tokens_in":2249,"tokens_out":813,"duration_ms":12014,"significance":"If the reported gains are causally attributable to dual-stream fast/slow decay partitioning under matched capacity and training budgets, the work would offer a practical multi-timescale temporal model for navigation that combines Mamba efficiency with reactive–strategic separation. Concrete multi-simulator evaluation and Unitree GO1 sim-to-real transfer are strengths for robotics impact. The contribution is primarily empirical architecture comparison rather than a new theoretical guarantee; significance therefore hinges on isolating the dual-stream design from capacity, hyperparameter, and variance confounds.","major_comments":[{"comment":"The central causal claim—that dual-stream (fast-decay reactive + slow-decay strategic) partitioning produces the reported gains (0.915 path efficiency; 576 median steps; 5% vs Vanilla-Mamba; lower timeouts)—cannot be assessed from the abstract alone. Parameter counts, FLOPs, matched model capacity, identical training budgets/hyperparameter schedules, and an ablation that removes only the dual-stream split (holding all else fixed) are not stated. Without these, the architectural contribution is not isolated from unmatched capacity or tuning (reader’s weakest assumption).","section":null},{"comment":"Reported metrics (576 median steps; 0.915 path efficiency; 5.0% improvement) lack error bars, number of seeds/runs, and statistical tests. For an empirical RL navigation comparison across LSTM, Transformer, and Vanilla-Mamba, variance and significance are load-bearing for ranking claims; their absence leaves the superiority statement under-supported on the available text.","section":null},{"comment":"Sim-to-real claims (Unitree GO1; static/dynamic obstacles; extended LiDAR ranges) are asserted without quantitative real-world metrics, failure modes, or matched sensing conditions versus baselines in the abstract. The full manuscript must supply these if the transfer claim is to support the dual-stream advantage under challenging sensing.","section":null}],"minor_comments":[{"comment":"Abstract-only review: section/equation/table citations cannot be grounded; all major concerns above must be re-checked against the full text (methods, tables, ablations).","section":null},{"comment":"Clarify definitions of path efficiency and timeout rate (normalization, goal distance, episode horizon) so the 0.915 and timeout comparisons are reproducible.","section":null},{"comment":"State explicitly how fast- and slow-decay rates (or Mamba timescale parameters) are chosen and whether they are fixed or learned.","section":null}],"recommendation":"uncertain","confidential_remarks":"This is an abstract-only review (full text unavailable). Recommendation is uncertain pending the full manuscript. If the full paper includes matched-capacity baselines, dual-stream-only ablations, multi-seed statistics, and quantitative GO1 results, the work may be suitable for minor or major revision rather than reject; if those controls are missing, major_revision or reject would be appropriate. Scope (cs.RO temporal RL for navigation) appears on-topic for a robotics venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing to know: StratMamba is a dual-stream Mamba for LiDAR obstacle avoidance—fast-decay reactive stream on high-frequency scans, slow-decay strategic stream for longer-horizon goals—and the abstract reports better path efficiency (0.915), fewer timeouts, and faster navigation (576 median steps, ~5% over vanilla Mamba) than LSTM, Transformer, and vanilla Mamba, with sim-to-real on a Unitree GO1.\n\nWhat is actually new is the application, not the motif. Multi-rate or hierarchical memory is familiar; wiring a fast/slow Mamba split specifically to reactive safety vs strategic planning for LiDAR nav is a legitimate architectural extension. Credit where it is due: concrete navigation metrics, multi-sim evaluation (IsaacLab and Gazebo), and a real quadruped demo with static and dynamic obstacles. The claim of more robust behavior at extended LiDAR ranges versus vanilla Mamba and Transformer is the part practitioners will care about. Circularity is low—this is empirical comparison against external baselines, not a derivation that collapses into a fitted target.\n\nThe soft spot is exactly the one the reader flagged, and it is proportionate, not fatal. From the abstract we cannot tell whether gains come from the dual-stream partition itself or from unmatched capacity, training budget, or variance. No parameter counts, FLOPs, matched hyperparameter schedules, ablation that removes only the split, or error bars/statistical tests are stated. Free parameters (decay rates, RL setup) are expected for this class of work; they just mean the causal story is still conditional. Nothing in the abstract contradicts the central claim or invents impossible numbers.\n\nThis is for people building temporal models for robot navigation and anyone comparing Mamba to Transformer/LSTM in RL. It does not open a new technology class, but it is a concrete system result with real-robot evidence. I would send it to a serious referee rather than desk-reject; ask specifically for capacity-matched baselines, an ablation on the dual-stream split, and variance. Worth a look in a robotics reading group if the full paper lands with those controls.","headline":"Useful dual-stream Mamba engineering for LiDAR nav with real GO1 transfer; abstract alone cannot pin the gains on the architecture split.","tokens_in":2934,"tokens_out":526,"would_cite":false,"duration_ms":11799,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A dual-stream Mamba model splits LiDAR memory into fast reactive and slow strategic paths for more efficient robot obstacle avoidance.","keywords":["StratMamba","Mamba","LiDAR","obstacle avoidance","dual-stream","robot navigation","sim-to-real","Unitree GO1"],"falsifier":"Retrain all baselines with matched parameter counts, identical training budgets, and multiple random seeds; if the dual-stream advantage in path efficiency and timeout rate disappears under those controls, the central architectural claim fails.","tokens_in":2938,"feed_emoji":"🤖","tokens_out":839,"duration_ms":8507,"temperature":0.7,"pith_summary":"StratMamba is a dual-stream temporal model built on Mamba that aims to solve a basic tension in LiDAR-based robot navigation: the robot must react instantly to nearby obstacles while still holding a long-horizon plan toward a goal. The architecture gives one stream a fast decay so it can process high-frequency LiDAR for reactive avoidance, and the other a slow decay so it can keep longer-horizon goal information for strategic planning. Evaluated in IsaacLab and Gazebo and transferred to a Unitree GO1 quadruped, the design reports the highest path efficiency (0.915), lower timeout rates, and the fastest median navigation (576 steps, 5 percent better than a single-stream Mamba) against LSTM, Transformer, and vanilla-Mamba baselines. Real-world tests further show more robust performance when the usable LiDAR range is extended, suggesting that the split itself helps balance safety and progress under imperfect sensing. If the dual-stream partition is the true source of the gains, the paper supplies a practical recipe for making state-space sequence models work better on mobile robots that must both react and plan.","feed_headline":"Dual-stream Mamba cuts robot path waste to 8.5 percent","feed_subtitle":"Fast reactive and slow strategic memory streams beat LSTM, Transformer, and vanilla Mamba on LiDAR avoidance.","key_machinery":"Dual-stream decay partitioning inside a Mamba state-space model: one stream with fast decay for reactive LiDAR processing and one with slow decay for strategic goal retention; the partition is what the authors claim supplies the temporal reasoning efficiency.","core_discovery":"A dual-stream Mamba temporal model that partitions memory into a fast-decay reactive stream for high-frequency LiDAR and a slow-decay strategic stream for goal information outperforms LSTM, Transformer, and single-stream Mamba baselines on LiDAR-based obstacle avoidance, achieving the highest path efficiency (0.915), lowest timeout rate, and fastest median navigation (576 steps) while transferring successfully from simulation to a Unitree GO1.","pith_inferences":["The same fast/slow decay partition may transfer to other high-frequency sensors such as event cameras or dense depth streams where reactive and strategic timescales also conflict.","If parameter-matched ablations confirm the gain, dual-stream decay could become a standard drop-in module for any Mamba-based navigation policy rather than a one-off architecture.","The lower timeout rate implies fewer recovery behaviors are needed, which could reduce energy and wear on physical platforms over long deployments."],"forward_implications":["Robots can keep both immediate collision-free reaction and longer-horizon goal progress without needing a separate planner or larger Transformer-style attention stack.","Path efficiency near 0.915 becomes a practical target for LiDAR-only obstacle avoidance under the dual-stream recipe.","Sim-to-real transfer on a Unitree GO1 is feasible for this style of dual-stream Mamba policy when static and dynamic obstacles are present.","Extended LiDAR ranges remain usable because the reactive stream can still suppress distant noise while the strategic stream holds the goal."],"fun_headline_variants":["Dual-stream Mamba hits 0.915 path efficiency on LiDAR avoidance","StratMamba reactive-slow streams beat LSTM Transformer vanilla Mamba","Fast-decay LiDAR plus slow goal memory cuts robot timeouts","Dual Mamba streams yield 576-step median paths on Unitree GO1","StratMamba partitions memory for path-efficient obstacle navigation"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That the reported gains come from the dual-stream decay split itself rather than from unmatched model capacity, hyperparameter choices, or training variance relative to the single-stream baselines.","fun_headline_variants_meta":{"raw":{"variants":["Dual-stream Mamba hits 0.915 path efficiency on LiDAR avoidance","StratMamba reactive-slow streams beat LSTM Transformer vanilla Mamba","Fast-decay LiDAR plus slow goal memory cuts robot timeouts","Dual Mamba streams yield 576-step median paths on Unitree GO1","StratMamba partitions memory for path-efficient obstacle navigation"]},"model":"grok-4.5","effort":"low","cost_usd":0.005848,"raw_usage":{"total_tokens":1544,"prompt_tokens":806,"num_sources_used":0,"completion_tokens":96,"cost_in_usd_ticks":58480000,"prompt_tokens_details":{"text_tokens":806,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":642,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":806,"tokens_out":96,"duration_ms":5311,"temperature":1.0,"reasoning_tokens":642,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-15T06:37:44.895563+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain all baselines with matched parameter counts, identical training budgets, and multiple random seeds; if the dual-stream advantage in path efficiency and timeout rate disappears under those controls, the central architectural claim fails.","supporting_citations":[],"review_version":1}