{"id":"2ae40242-f9c5-42e0-a45a-de9659d35bc9","arxiv_id":"2607.19306","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"OREN-Bubble* reconstructs a continuous signed distance field online and plans safe quadrotor trajectories through overlapping collision-free balls, demonstrating onboard flight in real time.","lead":"An autonomous drone system that builds a continuous signed-distance map of its surroundings and plans safe, low-collision-check flight paths through it, all onboard in real time. The map and the planner share one distance representation, which the authors argue makes cluttered-environment flight faster than conventional occupancy-map approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Bubble*'s Algorithm 1 as printed cannot expand the first bubble: line 20 chooses parents from OPEN∩B_c after line 6 removes c, so OPEN is empty on the first call; success and completeness rely on an unstated correction.","rationale":"The central claim is that OREN-Bubble* provides real-time planning with formal guarantees. I examined Algorithm 1 rather than the environmental-clearance assumption. The most load-bearing issue is the parent-selection rule: line 6 removes the expanded center from OPEN before line 20 searches for a parent in OPEN∩B_c, so the first expansion has no candidate node. The algorithm would loop once and return failure for every nontrivial query. The paper's own prose ('among nodes inside B_c') suggests line 20 should range over finite-g nodes in B_c, but neither the pseudocode nor the proof states this. This is not merely a typo in notation: Theorem 1, Theorem 2, and Corollary 1 all depend on successors being inserted with finite parent costs. Correcting the rule is likely simple, and the experimental system evidently uses such a correction, but the formal contribution as written is not self-contained. The reader's clear-path concern is real but secondary: even if a clear path exists, the printed algorithm fails. I therefore disagree that the clear-path assumption is the weakest point, though I keep the overall CONDITIONAL verdict because the issue is likely fixable and the experimental evidence suggests a working system.","tokens_in":24224,"tokens_out":7976,"duration_ms":81458,"concrete_test":"Re-implement Algorithm 1 verbatim, including OPEN←{p_s}, line 6's OPEN←OPEN\\{c}, and line 20's argmin over OPEN∩B_c. Run it on the Fig. 7 2D setup with any grid resolution and a nontrivial start–goal pair. If it returns failure after the first expansion, the pseudocode cannot support the reported results. Then change line 20 to choose k* among nodes in B_c with finite g-value (or keep c available as a parent) and rerun; if the planner succeeds and the proof still holds, the paper needs a corrected pseudocode and proof rather than a wholesale rejection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1's CalculateSuccessors (line 20) selects the parent of every newly reached boundary node as k* ∈ argmin_{k∈OPEN∩B_c} (g(k)+||j−k||_2). But the expanded center c is removed from OPEN on line 6 before B_c is constructed, and OPEN is initialized to {p_s} on line 2. On the first expansion, OPEN∩B_c is therefore empty. The argmin is undefined, no successor is inserted into OPEN, and the loop terminates with failure on line 17. Thus the algorithm as written returns failure for every nontrivial start–goal pair. The prose in Sec. 5.2 says the predecessor is chosen 'among nodes inside B_c,' not among OPEN nodes, so the pseudocode appears to conflict with the text and with the proof. Theorem 2's completeness proof assumes each boundary successor is inserted into OPEN with a finite parent cost; that cannot happen as printed. The delivered experiments and real flight must be using a different rule — for example, allowing all finite-cost nodes inside B_c, including c, to serve as parents. This is the most load-bearing point because termination, completeness, and failure detection all hinge on this step, independently of whether a 1.5Δ-clear grid path exists.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an integrated mapping and planning system for UAVs built around a non-truncated signed distance function (SDF). The mapping component, OREN, combines an explicit octree prior with an implicit neural residual to reconstruct a differentiable SDF online from depth point clouds. The planning component, Bubble*, searches over a graph of collision-free 'bubbles' grown from SDF clearance values, then uses MINCO to optimize a dynamically feasible trajectory within the resulting bubble corridor. The authors claim that OREN improves SDF estimation by 22% over baselines, that Bubble* finds trajectories spanning ~90 m in 1–3 s versus up to 10 s for baselines, and that the full pipeline runs onboard in real time on a Jetson Orin NX. Formal termination, completeness, and failure-detection guarantees are stated for Bubble*, and experiments include 2D and 3D simulated comparisons plus a real-world quadrotor flight.","tokens_in":24560,"tokens_out":4229,"duration_ms":44422,"significance":"If the central claims hold, this work makes a convincing case that mapping and planning should be co-designed around one distance representation: OREN provides the continuous clearance information that Bubble* consumes directly, and the planner's bubble graph unifies path search and safe-corridor construction. The formal completeness/failure-detection analysis, the code availability, and the real onboard demonstration are notable strengths. However, the evidence presented is thin in key places: the 3D planning results are single runs with no error bars, the 22% mapping improvement is inherited from the authors' prior conference paper rather than freshly evaluated here, and the printed search algorithm contains a load-bearing bug that prevents it from running as described. These issues do not invalidate the overall approach, but they must be addressed before the claims can be fully trusted.","major_comments":[{"comment":"The algorithm as printed cannot perform its first bubble expansion. On initialization OPEN={p_s}; the first iteration pops c=p_s and removes it from OPEN (line 6) before constructing B_c and calling CalculateSuccessors. Line 20 then takes argmin over k in OPEN∩B_c, which is empty on the first call. No successor receives a finite parent, no node is inserted into OPEN, and the loop exits on line 17 with failure. Thus the printed pseudocode returns failure for every nontrivial start–goal pair. The prose in Sec. 5.2 says the predecessor is selected 'among nodes inside B_c', which differs from the coded rule, and the completeness proof in Sec. 5.3 assumes each boundary node is inserted into OPEN with a finite parent cost, which cannot happen as written. This is load-bearing for termination, completeness, and failure detection. Please correct the pseudocode to match the intended rule — e.g., c","section":"Sec. 5.2, Algorithm 1, line 20"},{"comment":"The 3D planning comparison reports a single run per planner per environment, with no variance, seeds, or confidence intervals. The central efficiency claim — Bubble* has the lowest total planning time while matching path quality — rests on one trajectory. This is especially problematic for the stochastic RRT/RRT* baselines, where run-to-run variation can easily exceed the observed margins (e.g., Industrial: Bubble* 1008 ms vs A* 1228 ms; Forest: Bubble* 3776 ms vs A* 4152 ms). Please report multiple independent runs with standard deviations or interquartile ranges, and state the number of trials. Without this, the 3D planning conclusions are not quantitatively supported.","section":"Table 4"},{"comment":"The headline '22% SDF improvement' is not a fresh result of this manuscript. Table 2 is explicitly a summary of the authors' prior conference paper (Dai, Qian, et al., 2026), and the abstract and introduction present the 22% number as if it were newly demonstrated here. Since the current paper's contribution is the integrated OREN–Bubble* system, the mapping accuracy claim should either be re-evaluated in the paper's own simulated environments or clearly attributed as prior work in every appearance (abstract, introduction, and conclusion). As written, the 22% claim is effectively self-cited and does not provide independent evidence for the current system.","section":"Sec. 6.1 and abstract"},{"comment":"The completeness guarantee is conditional on the existence of a clear grid path with clearance at least 1.5Δ, and the text states that 1.5Δ is 'the smallest' clearance admissible under the grid-occupancy model without derivation. Mathematically, d(p_j)>Δ already ensures that the open bubble B(p_j,d(p_j)) contains the axis-aligned grid neighbors at distance Δ; the additional factor 1.5 is not justified in the proof. Furthermore, in unknown or partially mapped environments such a 1.5Δ-clear path may not exist even when a safe (nonzero-clearance) trajectory does, and the paper does not quantify how often Bubble* would report failure in navigable situations, nor provide a fallback. Please substantiate the 1.5Δ choice or revise the statement, and discuss the practical implications of this completeness condition.","section":"Sec. 5.3, Definition 1 and Theorem 2"}],"minor_comments":[{"comment":"Typographical and formatting issues: the title and abstract use 'UA Vs' with a space; Algorithm 1 uses the unusual arrow '←−' instead of '←'; the GitHub link in Data Availability contains a space ('erl oren bubble star demo') and should be a single URL.","section":"Throughout"},{"comment":"The simulated update-rate results (Fig. 6) report OREN running at 19.1–28.4 Hz, while the real-world experiment states OREN runs at 7 Hz. Please clarify whether this discrepancy is due to different hardware, point-cloud preprocessing, or parameter settings, and how the 30 Hz depth stream is handled at 7 Hz mapping rate.","section":"Sec. 6.2 vs Sec. 6.4"},{"comment":"The text says OREN achieves 'better or comparable' occupancy prediction, but in the Industrial environment OctoMap has a higher F1 (0.710 vs 0.688). This is not a major issue, but the claim should be phrased to reflect the actual pattern (higher recall, lower precision) rather than a blanket superiority.","section":"Table 3"},{"comment":"The disk-overlap construction in Eq. (14) is presented for two arbitrary overlapping balls, but the relationship to the specific bubble sequence returned by Bubble* (where consecutive bubble centers lie inside the previous bubble) is not formally connected. Also, the notation q_i is overloaded: Eq. (16) uses q_i for waypoints, while earlier q denotes a general point. Please align notation and make the overlap-construction validity conditions explicit.","section":"Sec. 5.4"},{"comment":"The related-work discussion of bubble-based planning mentions Ren et al. (2022) and the authors' prior work K.M.B. Lee et al. (2024), but the latter is not listed in the experimental comparisons. Given that Prior Bubble Cover is the closest algorithmic predecessor, a sentence explaining what Bubble* adds beyond it beyond the citation would help position the contribution.","section":"Sec. 2.2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript describes an attractive integrated system and provides useful formal analysis, but the printed Algorithm 1 is simply not executable for its first expansion — this is a correctness issue that must be fixed before the paper can be considered. The experimental evaluation also needs repeated trials and better handling of the self-cited 22% claim. The heavy reliance on the authors' own prior work is not disqualifying, but the current version overstates the novelty of the mapping results. With the pseudocode corrected and the experiments repeated, the paper could become a solid contribution; as is, major revision is warranted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the paper's integrated mapping-and-planning system is a plausible engineering advance, and the onboard flight demo is real evidence. But the stress-test is right: Algorithm 1 as printed cannot expand the first bubble. On line 6, c is removed from OPEN before CalculateSuccessors runs, and OPEN ∩ B_c is empty on the first call (OPEN was {p_s}). So the algorithm would return failure for every nontrivial start–goal pair. The prose in Sec. 5.2 says the predecessor is chosen 'among nodes inside B_c,' which is a different, workable rule; the pseudocode and proof follow the narrower OPEN∩B_c. That is load-bearing: termination, completeness, and failure detection all depend on successors entering OPEN. The delivered experiments must be using the prose rule or something equivalent. This needs a fix and a re-check of the guarantees before the results are taken as established.\n\nWhat's genuinely new: Bubble* as a graph-of-bubbles search with any-angle paths and formal claims, plus the integration with OREN's non-truncated SDF. The 2D search-efficiency result (91–99% fewer collision checks) is credible, and the real flight on a Jetson Orin NX is genuine. The public demo repo helps reproducibility.\n\nSoft spots: the 22% SDF improvement is carried over from the authors' own IROS paper, not evaluated here; Table 4 is single runs without error bars; the real-world demo is one trajectory. The completeness guarantee is conditional on a clear grid path with clearance ≥ 1.5Δ, and the paper does not quantify how often such a path exists or provide a fallback when the planner reports failure in a navigable space. These are real but not fatal—the planner's value is mostly empirical.\n\nWho this is for: people working on distance-field mapping and search-based planning for aerial robots. It deserves a serious referee; the bug is catchable and fixable, and the system is worth engaging with. My recommendation: send to review, but insist the authors correct Algorithm 1, add error bars or more runs, and clearly separate what is new here from the prior OREN results.","headline":"The integrated OREN-Bubble* system is real and the flight demo is genuine, but the Bubble* pseudocode as printed cannot expand the first bubble, so the completeness proof rests on an unstated fix; still worth refereeing.","tokens_in":25147,"tokens_out":2655,"would_cite":false,"duration_ms":27518,"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 shared signed-distance map can drive both mapping and planning, letting a quadrotor navigate unseen indoor spaces in real time.","keywords":["signed distance function","octree residual network","implicit neural representation","motion planning","bubble corridor","search-based planning","quadrotor autonomy","real-time mapping"],"falsifier":"In a grid of resolution Δ, build a straight corridor of width 2.0Δ (so centerline clearance is 1.0Δ, below 1.5Δ) and run Bubble* end-to-end; if it reports failure despite a centered straight trajectory being collision-free, the completeness precondition is shown to exclude a navigable environment.","tokens_in":24122,"feed_emoji":"🚁","tokens_out":8502,"duration_ms":71369,"temperature":0.7,"pith_summary":"This paper tries to establish that mapping and motion planning for autonomous flight belong in a single representation: a signed distance function (SDF), which returns the clearance to the nearest obstacle at every point. It introduces OREN, an online SDF reconstructor that combines a coarse explicit octree prior with a compact neural residual, and Bubble*, a search-based planner that treats each SDF query as a maximal collision-free ball and searches a graph of such balls. Together they let a quadrotor build a map and plan trajectories of about 90 meters in 1–3 seconds, issuing 91–99% fewer collision checks than a grid-based A* search. If the paper is right, distance information—not just occupancy—is the natural currency for both the map and the planner.","feed_headline":"Signed-distance bubbles let a drone plan 90-meter paths in seconds","feed_subtitle":"A hybrid neural-octree map supplies both obstacle distances and safe corridors, cutting collision checks by up to 99%.","key_machinery":"The bubble is the central object: for a grid node c, the SDF value d(c) defines an open ball B(c,d(c)) of guaranteed-free space. Bubble* expands these bubbles by adding boundary nodes—nodes with a grid neighbor outside the ball—as successors, so a single distance query replaces many per-cell collision checks and the search can jump across large free regions. OREN supplies the SDF: a semi-sparse octree (shallower layers kept dense so interpolation stays accurate away from surfaces) stores learnable distance and gradient values at vertices, gradient-augmented trilinear interpolation forms a coarse prior, and a small MLP decodes implicit vertex features into a residual correction, producing a c","core_discovery":"The central claim is that an accurate, non-truncated, differentiable SDF can be reconstructed online and then used as the sole geometric model for planning. OREN stores learnable SDF values and gradients at octree vertices, obtains a coarse prior by gradient-augmented interpolation, and decodes a neural residual from implicit features to recover fine surface detail; an occupancy decoder supervises the sign of the prediction. Bubble* queries this SDF at grid nodes, grows an open ball of radius equal to the signed distance at each node, and expands a graph whose edges connect nodes inside a bubble to nodes on its boundary, so one distance query validates an entire region. The paper proves term","pith_inferences":["A testable extension is to lower the 1.5Δ clearance precondition: in a corridor whose width is between 2Δ and 3Δ, a centerline path may be safe for a small vehicle yet fail the clear-grid-path condition, so Bubble* would report failure; a resolution-adaptive or hybrid fallback would close this gap.","The planner only needs a distance oracle, so the same bubble-graph idea could be applied to configuration-space distance fields for manipulators or to learned distance fields from other sensors, not just OREN.","The safety margin r in the trajectory optimization depends on the accuracy of far-field SDF values; if OREN's error grows with distance in unseen areas, the effective clearance shrinks, suggesting that an online uncertainty estimate for bubbles would strengthen the guarantees.","Because multiple Bubble* plans can be concatenated into one overlapping-bubble sequence, the method supports waypoint-chained missions without re-planning from scratch, which the paper demonstrates with a looping three-segment trajectory."],"forward_implications":["A planner that consumes signed distance directly can replace per-cell collision checks with one distance query per bubble, reducing collision checks by 91–99% in the paper's experiments.","Because the bubble corridor is produced by the search itself, path-finding and safe-corridor construction cease to be separate stages, which keeps trajectory optimization fast (e.g., 523 ms vs 1005 ms in the industrial test).","Non-truncated SDF support means distance queries far from surfaces remain valid, which is what lets global planning reason about clearance across large spaces rather than only near surfaces.","The termination, completeness, and failure-detection guarantees mean the planner can act as a decision module: it returns a corridor when one with sufficient clearance exists and reports failure otherwise.","If OREN's mapping accuracy transfers, the same loop can run on compute-limited platforms, since the full mapping-planning stack ran onboard in real time in the paper's flight tests."],"fun_headline_variants":["Neural-octree SDF powers 90-meter drone paths in seconds","Bubble planner uses SDF to cut collision checks by 99%","Drone builds SDF map in real time, plans 90-meter trajectories","Hybrid SDF map and bubble graph plan 90-meter drone paths"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The completeness guarantee only holds when a 'clear grid path' exists—a grid path where every node has signed distance at least 1.5 times the grid resolution (Definition 1, Theorem 2); in unknown or narrow environments such a path may not exist even when a safe trajectory does, and the paper does not quantify how often this happens or provide a fallback.","fun_headline_variants_meta":{"raw":{"variants":["Neural-octree SDF powers 90-meter drone paths in seconds","Bubble planner uses SDF to cut collision checks by 99%","Drone builds SDF map in real time, plans 90-meter trajectories","Hybrid SDF map and bubble graph plan 90-meter drone paths"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00102,"raw_usage":{"total_tokens":4183,"prompt_tokens":830,"completion_tokens":3353,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":574,"completion_tokens_details":{"reasoning_tokens":3272}},"tokens_in":574,"tokens_out":3353,"duration_ms":21068,"temperature":1.0,"reasoning_tokens":3272,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T12:47:59.094859+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a grid of resolution Δ, build a straight corridor of width 2.0Δ (so centerline clearance is 1.0Δ, below 1.5Δ) and run Bubble* end-to-end; if it reports failure despite a centered straight trajectory being collision-free, the completeness precondition is shown to exclude a navigable environment.","supporting_citations":[],"review_version":1}