{"id":"7c0554a9-e663-40ca-83aa-a5b5335b3613","arxiv_id":"2507.21517","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A floor-stair topological decomposition lets a trained 2D attention-based exploration policy guide a robot through multi-floor indoor environments.","lead":"LITE is an exploration system that splits a multi-floor building into a network of floors and staircases, then uses a learned attention-based policy to explore each floor and a state machine to move between floors. The authors show that this decomposition lets existing 2D exploration methods work in 3D multi-floor buildings, and they test it in simulation and on a real quadruped robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stair-detector generalization is the single point of failure: no held-out detection evaluation on HM3D or real scenes means the multi-floor claim is unverified.","rationale":"The reader identified the same load-bearing assumption I would: the stair detector trained on internet images is the single point of failure for the multi-floor exploration claim. A missed stair prevents the floor-stair topology edge from being created, so the FSM can never leave the current floor; a false stair can send the agent off-course. The paper reports no stair-detection performance on HM3D or the real environment, and the multi-floor experiments provide no per-scene transition success or detection recall. This is not a reason to reject the paper, because the 2D-policy contribution is independently supported by comparative results and ablations, and the topological framework is plausible. But it is a reason to keep the verdict conditional: the multi-floor central claim should be verified with a held-out stair-detection evaluation before being accepted as established. The proposed test would settle whether the concern actually lands, and the reader's conditional recommendation remains appropriate.","tokens_in":11998,"tokens_out":3691,"duration_ms":51212,"concrete_test":"Have the authors release the stair-detector weights and evaluate detection recall/precision on the six HM3D multi-floor scenes used in §V-C (or a held-out set of HM3D stairways), computing per-scene stair instance recall and the corresponding rate of successful FSM floor transitions over multiple episodes. If every required stair edge is detected and floor transitions succeed in all scenes, the concern is resolved; if any scene misses a required stair, the multi-floor results in Table III are not attributable to LITE in the claimed generality.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The multi-floor central claim depends entirely on the YOLO11 stair detector. The detector is trained on 1,511 self-labeled internet images and reported only on that dataset (mAP 0.954, §IV-D-1), then 'directly deployed' in §V-C and §V-D with no detection evaluation on the HM3D scenes or the real teaching building. This is not an internal inconsistency, but it is an unvalidated empirical generalization at the system's single point of failure. In Algorithm 1, stair instances detected at lines 4-5 are the only mechanism that creates topology edges (lines 11-13); in Algorithm 2, the transitions GoingToStair → OnStair → ExploringFloor and the termination condition AllExplored all depend on detected stair edges and eon_stair. A missed stair means the edge is never created and the FSM never leaves the current floor; a false positive can send the agent toward a nonexistent transition. Section V-C reports only aggregate CR and steps over six scenes, with no per-scene floor-transition success or detection recall, and Section V-D is qualitative. The 2D-policy experiments are comparatively well supported, with error bars and ablations, but the multi-floor contribution remains conditional on a detector whose generalization is asserted rather than measured.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents LITE, a multi-floor exploration system that combines a floor-stair topological graph, a YOLO11-based stair detector, and an attention-based 2D exploration policy trained with PPO. The 2D policy selects global goals from an 8-channel stacked map; a finite state machine uses detected stairs to transition between floors. Experiments compare LITE-2D to frontier, sampling, and neural baselines on HM3D and MP3D, ablate components, test on six HM3D multi-floor scenes, and demonstrate a real-world quadruped deployment.","tokens_in":12226,"tokens_out":4450,"duration_ms":47732,"significance":"If the claims hold, the main significance is a practical recipe for lifting 2D exploration policies to multi-floor buildings without 3D RL, plus evidence that attention-based goal selection improves over prior global-goal policies. The 2D experiments include standard deviations and three ablations, and the trained policy generalizes to held-out MP3D scenes; a real-world validation is a useful bonus. The main weaknesses are that the multi-floor result hinges on an unvalidated stair detector and that the reported gains over ANS Global are small and not statistically supported, so the central 'significantly outperforms' claim is stronger than the evidence.","major_comments":[{"comment":"The multi-floor architecture depends on the YOLO11 stair detector as the sole mechanism for creating topology edges and triggering floor transitions, but its performance on HM3D and real-world data is never evaluated. The detector is trained on 1,511 self-labeled internet images and its reported mAP (0.954) is only on that dataset; in §V-C it is 'directly deployed' on six HM3D scenes, and in §V-D on a teaching building, with no detection recall/precision on those data. A missed stair prevents the edge from being created (Algorithm 1 lines 4-5 and 11-13), so the FSM never leaves the floor; a false positive can send the agent toward a nonexistent transition. This is not an internal inconsistency, but it is an unvalidated empirical generalization at the system's single point of failure. Please report per-scene stair detection performance on the test scenes and, ideally, the number of successful floor transitions for each scene; without this, the multi-floor claim is unverified.","section":"Algorithms 1 and 2, §IV-D-1, §V-C"},{"comment":"The claim that LITE-2D \"significantly outperforms\" ANS Global is not supported statistically. The reported gains are CR +1.7% and SR +6.6%, and the standard deviations overlap (0.918 ± 0.071 vs 0.901 ± 0.097). No significance tests or per-scene paired comparisons are provided. Please add paired statistical tests (e.g., paired bootstrap or Wilcoxon signed-rank) and/or per-scene results to substantiate the claimed significance.","section":"§V-B, Table I"},{"comment":"The multi-floor comparison reports only aggregate mean CR and steps over six scenes, without variance, per-scene results, or any measure of stair detection/transition success. With n = 6 scenes and no variance, the claim that LITE-2D outperforms other methods in multi-floor exploration is not established. Please provide a per-scene table with CR, steps, and floor-transition outcomes (e.g., number of successful stair transitions, detection recall per scene).","section":"Table III"},{"comment":"The frontier guidance reward in Eq. (2) embeds the frontier-selection heuristic into training. Since the comparison in Table I includes frontier-based baselines, the reported improvements may partly reflect the incorporation of frontier information into the reward rather than a fundamentally superior exploration policy. The ablation in Table II shows that the frontier reward helps, but the comparison to frontier baselines should be framed accordingly, or the policy should be compared to baselines trained with the same frontier signal. This does not invalidate the method, but it tempers the \"significantly outperforms all baselines\" claim.","section":"§IV-C-4, Table I"}],"minor_comments":[{"comment":"The term \"Learning-Intergrated\" in the Introduction should be \"Learning-Integrated\"; the conclusion repeats the typo as \"leaning-integrated\".","section":"Abstract/Introduction"},{"comment":"The first sentence of §V-C, \"We further perform experiments in 6 multi-floor indoor environments of HM3D to c the ability,\" is incomplete; \"to c\" should likely be \"to test\".","section":"§V-C"},{"comment":"The statement that LITE \"constructs the two-floor environment to a three-floor topology\" is confusing; if the real building has two explored floors and two stair edges, please describe the topology precisely.","section":"§V-D"},{"comment":"The ablation results report no standard deviations, while Table I includes them; please add standard deviations to Table II for consistency and to assess variability.","section":"Table II"},{"comment":"The notation A(at|st, πθ) in Problem 1 is unclear; it should be clarified that A is the covered area at time t given the action, state, and policy parameters.","section":"Eq. (1)"},{"comment":"The label \"Frontier TrainingGudance\" contains a typo, and the orientation embedding is not clearly defined in the text; please provide a concise definition.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a robotics journal. My main concern is the gap between the strong, well-supported 2D results and the unvalidated multi-floor claims, which rest entirely on a stair detector whose generalization is asserted rather than measured. The missing detector evaluation and statistical support are load-bearing; I would recommend asking for these additions before acceptance. I do not see grounds for rejection if the missing evidence can be provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version. LITE is a system that decomposes multi-floor buildings into a floor-stair graph and then drops a 2D exploration policy onto each floor, with a finite state machine to handle stair transitions. That combination is genuinely new — no one else has integrated an RL-based 2D explorer into multi-floor exploration this way. The 2D policy itself is an attention-based goal selector in the ARiADNE/ANS family, with a frontier-guided reward and an orientation embedding. The 2D experiments are the strongest part: they report standard deviations, run ablations showing each component contributes, and generalize to MP3D. That part is honest and reasonably solid.\n\nThe soft spot is exactly where the stress-test landed. The whole multi-floor architecture depends on the YOLO11 stair detector. It is trained on 1,511 self-labeled internet images and reported only on that dataset. Then it is 'directly deployed' on HM3D scenes and a real building, with no detection evaluation on those targets. In Algorithm 1, a missed stair means the edge never gets created and the FSM never leaves the floor. The paper does not give per-scene floor transitions or detection recall. That is a load-bearing unmeasured assumption, not a minor omission.\n\nOther issues are more minor. The gains over ANS Global in Table I are +1.7% CR and +6.6% SR, called 'significantly better' without significance tests. The multi-floor comparison in Table III has six scenes and no variance. No code or data are released, so reproducing the 2D results would take real effort. The real-world demo is qualitative but still useful supporting evidence.\n\nIn short: the 2D policy is probably fine, and the topological integration idea is worth taking seriously. But the paper's central multi-floor claim is conditional on a detector whose generalization remains unmeasured. That needs to be fixed before publication, either by evaluating detection on the target scenes or by opening the loop so missed stairs are recoverable. I'd send it to review — the idea deserves referee time — but the reviewers should push on this exact point.","headline":"A useful multi-floor exploration system where the 2D policy work is solid but the stair detector—the load-bearing component—has unmeasured generalization, making the central claim conditional.","tokens_in":12768,"tokens_out":2025,"would_cite":true,"duration_ms":22049,"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":"A floor-stair graph lets 2D explorers cover multi-floor buildings.","keywords":["multi-floor exploration","floor-stair topology","attention-based exploration policy","deep reinforcement learning","stair instance segmentation","autonomous exploration","topological mapping","quadruped robot"],"falsifier":"Measure the stair detector's precision and recall against hand-labeled stair masks in the same simulated multi-floor scenes used for evaluation and in a set of real staircases with varied lighting. If any nontrivial fraction of stairways is missed, the finite state machine never leaves ExploringFloor and the agent cannot complete the second floor.","tokens_in":11773,"feed_emoji":"🤖","tokens_out":9068,"duration_ms":89985,"temperature":0.7,"pith_summary":"The paper proposes LITE, a system that represents a multi-floor building as a floor-stair topology: a graph whose nodes are 2D floor maps and whose edges are detected stairways. The central claim is that this decomposition is enough to turn any 2D exploration method, learning-based or classical, into a 3D multi-floor explorer, because the rich information in an indoor scene is dense inside each floor and sparse between floors. The paper further claims that its attention-based 2D exploration policy, trained with PPO and frontier guidance, chooses global goals more efficiently than the compared baselines, and that the complete system transfers to a real quadruped robot in a four-floor teaching building. If these claims hold, multi-floor exploration no longer requires 3D-specific planners or high-dimensional 3D observations.","feed_headline":"Stair detection lifts 2D exploration to multi-floor buildings","feed_subtitle":"Attention-based goals raise coverage and success over baselines, and the system transfers to a real quadruped robot.","key_machinery":"The load-bearing object is the floor-stair topology graph $G=(V,E)$, where each node is a floor's 2D grid map and each edge is a stairway, maintained online by a YOLO11-based instance segmentation model. A finite state machine over four states (ExploringFloor, GoingToStair, OnStair, AllExplored) decides when the current floor is done, directs the agent to the nearest stair center, guides it through the stair, and stops when no new stair edges appear. The 2D policy that runs inside this graph is an attention-based explorer: a CNN extracts features from an $8 \\times L \\times L$ stacked local-and-global map, a transformer encoder with positional embeddings captures long-range spatial dependencies, and a goal-generation head emits a global goal, while a frontier-distance reward biases training toward informative regions. The Fast Marching Method supplies local paths to each global goal.","core_discovery":"The discovery is a representational reduction: by treating stairs, not geometry, as the only inter-floor coupling, LITE reduces multi-floor exploration to repeated instances of 2D exploration plus a finite state machine. On single-floor benchmarks, LITE-2D reports higher coverage ratio (0.918 vs. 0.901), higher coverage area, higher area weighted by path length, and higher success rate (0.473 vs. 0.407) than the strongest baseline, with ablations attributing the gains to the attention mechanism, the frontier guidance reward, and the orientation embedding. In six multi-floor scenes, the framework averages 0.929 coverage ratio and finishes in 1708 steps, and every baseline explorer placed inside the same topology completes multi-floor exploration. A real-world trial on a quadruped robot in a four-floor teaching building builds a three-floor topology while exploring the first two floors, supporting the generalization claim.","pith_inferences":["The same finite state machine should transfer to elevators, ramps, and escalators once the semantic detector is retrained for those transition types; the paper only demonstrates stairs.","The paper's stair-detection accuracy (0.954 mAP 50-95) is reported only on the detector's own internet-image dataset, so the robustness of floor transitions inside the simulated and real scenes is not directly measured; a held-out stair-mask evaluation on those scenes would separate exploration-policy gains from perception luck.","Because the finite state machine creates a topology edge only when a stair is detected, a missed stair freezes the agent on the current floor; the system's sensitivity to detector threshold choices is a natural stress test."],"forward_implications":["If the central claim is right, a 2D reinforcement-learning explorer can be deployed in multi-floor buildings without any 3D training data or 3D observation space.","The topology is method-agnostic: frontier-based and sampling-based explorers placed in the same graph also complete multi-floor exploration in the paper's six test scenes, though with lower efficiency and more steps.","The attention mechanism is doing real work: removing it drops coverage ratio from 0.918 to 0.902 and success rate from 0.473 to 0.380, close to the strongest baseline without attention.","Frontier guidance is what keeps the agent from skipping hard-to-reach regions; its removal lowers area-per-path-length and success rate more than coverage ratio.","The hardware gap can be bridged: the real quadruped used a different sensor suite and controller than training, and the stair model transferred without fine-tuning."],"supporting_citations":[{"why":"Supplies the frontier-based exploration baseline and the greedy nearest-frontier goal rule that LITE-2D outperforms.","marker":"[4]"},{"why":"Supplies the sampling-based next-best-view baseline whose goals are transferred into LITE's framework for comparison.","marker":"[5]"},{"why":"Supplies the active neural SLAM baseline and the task formulation for 2D exploration used in the paper.","marker":"[9]"},{"why":"Supplies the YOLO11 instance segmentation architecture used for real-time stair detection and topology building.","marker":"[12]"},{"why":"Supplies the PPO algorithm used to train the attention-based exploration policy.","marker":"[24]"},{"why":"Supplies the Fast Marching Method used as the local planner that navigates to global goals.","marker":"[25]"},{"why":"Supplies the vision transformer paradigm that the attention encoder is built on.","marker":"[26]"},{"why":"Supplies the segmentation model used for semi-automatic annotation of the 1,511 stair images used to train the detector.","marker":"[27]"},{"why":"Supplies the simulated indoor scenes used for training the 2D policy and evaluating multi-floor exploration.","marker":"[29]"},{"why":"Supplies the additional large indoor scenes used to test the policy's generalization without retraining.","marker":"[30]"}],"fun_headline_variants":["Stair-based topology reduces multi-floor to repeated 2D","Attention-driven explorer climbs floors via stair topology","LITE: multi-floor exploration via stair-aware 2D policy","Topological stair model lifts 2D exploration to 3D","Multi-floor exploration simplified by stair decomposition"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the stair detector, trained on 1,511 self-labeled internet images, recognizes stairs in the unseen simulated scenes and the real teaching building without scene-specific fine-tuning; every transition between floors depends on a successful detection.","fun_headline_variants_meta":{"raw":{"variants":["Stair-based topology reduces multi-floor to repeated 2D","Attention-driven explorer climbs floors via stair topology","LITE: multi-floor exploration via stair-aware 2D policy","Topological stair model lifts 2D exploration to 3D","Multi-floor exploration simplified by stair decomposition"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000256,"raw_usage":{"total_tokens":1586,"prompt_tokens":969,"completion_tokens":617,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":585,"completion_tokens_details":{"reasoning_tokens":538}},"tokens_in":585,"tokens_out":617,"duration_ms":7325,"temperature":1.0,"reasoning_tokens":538,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T12:39:38.728034+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the stair detector's precision and recall against hand-labeled stair masks in the same simulated multi-floor scenes used for evaluation and in a set of real staircases with varied lighting. If any nontrivial fraction of stairways is missed, the finite state machine never leaves ExploringFloor and the agent cannot complete the second floor.","supporting_citations":[],"review_version":1}