{"id":"659de11e-815f-488b-a7e6-96ac80440d15","arxiv_id":"2505.17438","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A three-module mapping, search, and trajectory optimization system lets a LiDAR drone avoid obstacles at up to 15 m/s in dense simulated forests and over 11 m/s outdoors.","lead":"This paper presents HEPP, a complete drone navigation system that maps, searches, and plans at speeds up to 15 m/s in cluttered spaces with only a few milliseconds of computation per cycle. It matters because it pushes autonomous drones closer to practical high-speed flight in dense environments, where prior systems typically fail or slow down.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RESDF's nearest-point distance can overestimate true obstacle clearance; without conservative inflation or unknown-space handling, the safety claim of 15 m/s flight is not guaranteed.","rationale":"The reader's weakest assumption identifies exactly the load-bearing weakness: RESDF substitutes distance-to-nearest-point for true distance-to-obstacle, and the paper offers no conservative inflation or completeness guarantee. My analysis sharpens the mechanism: the quantization in Eq. (1) introduces an upper-bound overestimate of up to sqrt(3)r/2, while the binary occupancy test in Eq. (8) classifies unobserved space as free, making the overestimate unbounded. This is not merely a formal nicety; the collision cost in Eq. (24) is zero whenever D exceeds the threshold ds, so the optimizer can certify a trajectory as safe while the true clearance is negative. Every module of the system relies on this field: the path search's visibility condition, the trajectory initialization along a 'collision-free' path, and the final gradient-based refinement. If this assumption fails, the system can plan directly into an obstacle, which would falsify the abstract's claim of safe high-speed obstacle avoidance. I agree with the reader's CONDITIONAL verdict because the concern is about an unproven safety guarantee rather than a demonstrated failure in the reported experiments; the simulated and real tests happened to work, but the paper gives no reason to expect safety in the general cluttered environments it claims to handle. A concrete ground-truth clearance audit and a thin-wire test would settle whether the concern lands. Other issues, such as the unsupported 'close to global optimum' claim and the cherry-picked timing comparison, are real but secondary: they affect trajectory quality and presentation, not the core safety argument.","tokens_in":19068,"tokens_out":7634,"duration_ms":67890,"concrete_test":"In the simulated column-and-circle forest at vlim=15 m/s, after each successful HEPP run, compute the true minimum distance from the executed trajectory to the ground-truth obstacle primitives (not the point cloud). If any trajectory point has true clearance less than the vehicle radius (or less than ds - sqrt(3)r/2, the voxel-center quantization bound), then the RESDF safety check failed. A second, targeted test: add a 2 cm diameter wire perpendicular to the flight path and sample it with the same 0.1 m downsampling as the rest of the map, so it yields at most one or two points per meter; run HEPP at 15 m/s and check whether the planner flies through the wire. If it does, the unknown-space-as-free behavior is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central safety mechanism is the Robocentric ESDF defined in Section III-D. Eq. (9) sets D(wxr) = ||K(wxr,1) - wxr||, the distance to the single nearest point in the sparse, sliding point cloud, and Eq. (8) classifies any position without a point within radius r as Free. Two consequences make this unsafe as the sole collision check. First, since points are quantized to voxel centers by Eq. (1), D overestimates the true distance to the obstacle surface by up to sqrt(3)r/2, and this overestimate is unbounded when a surface is undersampled or occluded. Second, and more seriously, the map contains only observed surfaces; space that is unknown because it is outside the LiDAR FOV, shadowed by other obstacles, or too thin to return a point is treated as free by the occupancy check and by the collision cost. In Eq. (24), Gs = ds - D, and the penalty in Eq. (22) is zero whenever D > ds, so the optimizer receives no repulsive signal in any region where the nearest sampled point is far away, regardless of whether an unsampled obstacle actually occupies that region. The paper provides no completeness or density guarantee, no conservative inflation, and no explicit unknown-space penalty. Because every trajectory is pushed away only by this nearest-point field, the central claim of safe 15 m/s obstacle avoidance in cluttered environments rests on the unstated assumption that the point cloud is a dense, complete sample of every obstacle surface in the replanning horizon. That assumption is not established and is violated in realistic thin-structure or occlusion scenarios.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents HEPP, a complete UAV navigation stack for high-speed obstacle avoidance consisting of three modules: an incremental robocentric point-cloud map with direct nearest-neighbor distance and gradient queries (IROP-Map/RESDF), an obstacle-aware topological path search that expands multiple homotopy candidates on a visual plane, and a MINCO-based trajectory optimizer with adaptive time allocation. The authors report sub-10 ms end-to-end latency, an 89.5% reduction in mapping time, a 79.24% reduction in total time relative to existing methods at 15 m/s, and successful autonomous flights in simulation and real-world cluttered environments at speeds up to 13 m/s speed limits. The abstract and contributions also claim that the planned trajectories are close to the global optimum in both temporal and spatial domains.","tokens_in":19397,"tokens_out":7734,"duration_ms":64087,"significance":"The paper is a solid systems contribution: the incremental mapping data structure, the topological path search, and the integrated full-stack validation are clear strengths, and the real-world flights at 8-11 m/s demonstrate that the system works in practice. The reported mapping and path-search efficiency gains are plausible and well motivated, and the paper is generally careful in describing the algorithmic pipeline. However, the headline claims -- safe 15 m/s obstacle avoidance, a 79.24% latency reduction over existing methods, and closeness to global optimality -- are not fully supported by the evidence as presented, because the collision model is not conservative, the latency comparison is not like-for-like, and the optimality claim is not operationalized. These issues are load-bearing for the paper's central claims, so the contribution is significant but conditional on revision.","major_comments":[{"comment":"The RESDF collision model is not conservative. Eq. (9) defines D(wxr) as the distance to the single nearest map point, and Eq. (8) labels any position with no point within radius r as Free. Because points are quantized to voxel centers by Eq. (1), D can overestimate true clearance by up to sqrt(3)r/2 at observed surfaces, and the error is unbounded wherever a surface is undersampled or outside the sensor field of view. The collision cost in Eq. (24) is zero whenever D > ds, so unobserved or gappy space produces no repulsive signal at all. The paper provides no completeness bound, no conservative inflation, and no unknown-space penalty, so the 15 m/s safety claim rests on the unstated assumption that the point cloud densely samples every obstacle surface in the planning horizon. Please add a conservative treatment or explicitly qualify the safety claim as empirical.","section":"Section III-D, Eqs. (9)-(10), (24)"},{"comment":"The headline latency comparison is not like-for-like. The 79.24% figure in the abstract is (26.30 - 5.46)/26.30, which uses only Ours and Agile-Auto from the vlim = 15 m/s block of Table III. Agile-Auto is a learning-based end-to-end policy that consumes depth images and has no mapping, path-search, or trajectory-optimization stages (those entries are shown as '/'), so comparing total algorithm latency conflates different system architectures. Moreover, Ego-PlannerV2 and Fast-Dodge, which the text says failed at 10 and 15 m/s, are omitted from the table; showing them with failed or zero success rates would make the comparison transparent. Please revise the claim to name the specific baseline and report all baselines in the table.","section":"Table III and Abstract"},{"comment":"The global-optimality claim is not operationalized. The abstract and Section I state that the planned trajectory is close to the global optimum in temporal and spatial domains, but the paper never defines an optimality metric or computes an optimality gap. Selecting the trajectory with the smallest cost J in Eq. (20) among K = 4 angular samples and a finite set of topologies only guarantees best-of-finite-set with respect to the same objective being optimized; it does not establish closeness to the true global optimum of the nonconvex problem. Please either remove the global-optimality wording or support it with a concrete optimality bound or benchmark on tractable instances.","section":"Abstract and Contributions"},{"comment":"Eq. (18) appears to be an incorrect feasibility condition when v0 > 0. For a constant-acceleration phase from v0 to vd followed by deceleration to zero, the required distance is (vd^2 - v0^2)/(2ad) + vd^2/(2ad) = (2vd^2 - v0^2)/(2ad), not ((vd - v0)^2 + vd^2)/(2ad). The published expression is smaller when v0 > 0, so the check in Algorithm 2 can accept an infeasible initial time allocation during in-flight replanning, where v0 is the current velocity. Please correct Eq. (18) or state explicitly that v0 is assumed to be zero.","section":"Section V-B, Eq. (18)"}],"minor_comments":[{"comment":"There is a typo: 'Meanwile' should be 'Meanwhile'. Also, the sentence 'The red font indicates the case of error' does not match Table III, where no red font is visible; either add the highlighting or remove the sentence.","section":"Section VI-D"},{"comment":"The condition 'if ι == Default' is confusing. Please clarify what Default represents and distinguish explicitly between 'no occlusion', 'first encounter with a new obstacle', and 're-encounter with a previously labeled obstacle'.","section":"Algorithm 1, Lines 10-16"},{"comment":"The deletion rule is described as removing points when IΔ < 0, but Eq. (6) keeps points with IΔ > 0; the sign convention and the treatment of IΔ = 0 should be stated unambiguously.","section":"Section III-B, Eqs. (5)-(6)"},{"comment":"Please specify whether the reported latencies are means over successful runs only, and include standard deviations or per-run variability; this is important for the high-speed claim given the modest number of runs (20 per group).","section":"Table III"},{"comment":"The limitations paragraph discusses trajectory selection but does not mention the RESDF unknown-space assumption or the lack of a formal safety guarantee; adding this limitation would make the scope of the safety claim clearer.","section":"Section VI-D and Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent systems contribution, and the efficiency results for the mapping and topological-search modules are likely useful to the community. The main risk is overclaiming: the 15 m/s safety guarantee, the 79.24% latency comparison, and the global-optimality statement are stronger than the evidence supports. These are fixable within the manuscript's scope by conservative treatment or explicit qualifications, by reporting all baselines, and by removing or bounding the optimality claim; I therefore recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real engineering system, not a toy, and the core ideas are worth taking seriously. The system is coherent, the modules fit together, and the measured latencies are impressively low. The robocentric i-Octree map with direct nearest-neighbor distance/gradient queries is a sensible replacement for explicit ESDF maintenance, and the visual-plane topological search is a clever way to get multiple distinct paths quickly. The adaptive MINCO time allocation also makes sense as a way to avoid bad local minima at high speed. I believe the system works in the demonstrated regimes.\n\nThe soft spots are real but mostly at the level of claims, not architecture. The biggest one is safety. RESDF is a nearest-neighbor distance in a sparse, sliding, voxel-quantized point cloud, and Eq. (8) treats any empty neighborhood as free. Eqs. (22)-(24) only penalize when the nearest observed point is closer than d_s, so unknown or occluded space—outside the LiDAR FOV, behind a trunk, or on a thin pole—gets no repulsive signal. In a dense forest of thick trunks this is likely fine; that is what the experiments show. But the paper says \"guarantee safety\" and \"close to the global optimum,\" and neither follows from the model. The stress-test note about sqrt(3)r/2 quantization bias is less important because a safety threshold d_s > r absorbs that. The unknown-space problem is the real one, and it should be stated as an assumption or handled with a conservative unknown-cost/inflation.\n\nThe evaluation has the usual second-system problems. Table III silently drops Ego-PlannerV2 and Fast-Dodge at 10 and 15 m/s; the text says they failed, which is honest, but the table should show it. The 79.24% number is taken from the 15 m/s row against Agile-Auto, a learning method with different sensors and hardware, so it should not be generalized to \"existing methods.\" There are no error bars, no code release, and the real flights are qualitative. The final limitation paragraph about cost-based trajectory selection is a good sign; the authors know the selection rule is not fully optimal.\n\nCitation pattern is broadly appropriate: prior work such as ROG-Map, i-Octree, and MINCO is named and the comparisons are mostly fair, aside from the omission issue.\n\nVerdict: referee it. The integration is novel, the empirical work—even if incomplete—is far beyond a toy, and the safety/optimality concerns can be addressed with language and a few added experiments. I would not desk-reject; I would send it back for major revision that fixes the claims and the missing failure rows.","headline":"A credible, well-engineered high-speed planner whose empirical results are promising, but whose safety and near-optimality claims outrun the evidence; referee it, but require caveats and more rigorous evaluation.","tokens_in":19991,"tokens_out":4140,"would_cite":true,"duration_ms":36931,"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":"HEPP claims a complete perception-and-planning pipeline that lets a drone avoid obstacles at 15 m/s in dense, unknown environments with planning loops of a few milliseconds, taking 79.24% less time than existing methods.","keywords":["high-speed UAV navigation","obstacle avoidance","incremental robocentric mapping","Euclidean signed distance field","topological path search","trajectory optimization","time allocation","lidar-based perception"],"falsifier":"Run the pipeline in a cluttered forest while computing a ground-truth signed distance field from dense registered scans; if any trajectory that the planner accepts has true clearance below the safety threshold $d_s$, or if the mean error between RESDF and the ground-truth distance at sampled trajectory points is comparable to $d_s$, the nearest-neighbor distance assumption is falsified.","tokens_in":18827,"feed_emoji":"🚁","tokens_out":6952,"duration_ms":51891,"temperature":0.7,"pith_summary":"HEPP sets out to prove that a drone can avoid obstacles at 15 m/s in dense, unknown spaces by making every computation local: a moving point-cloud map replaces global occupancy grids and distance fields, a topological search produces several genuinely different paths, and a gradient-based optimizer with adaptive time allocation turns each path into a trajectory. The paper reports that this pipeline runs in a few milliseconds per iteration, taking 79.24% less time than existing methods at high speed, and that picking the lowest-cost trajectory among the candidates brings the flight close to the global optimum in both duration and distance. A reader should care because high-speed autonomy in clutter has previously been limited to sparse fields or learning-based systems with narrow training envelopes; this work claims a fully classical pipeline that is both fast enough and robust enough for real forest flights.","feed_headline":"New system cuts drone replanning time by 79% at 15 m/s","feed_subtitle":"A mapping, path-search, and trajectory trio keeps quadrotors near-optimal in cluttered, unknown spaces.","key_machinery":"The load-bearing object is RESDF, the Robocentric Euclidean Signed Distance Field: for any query position, the distance to the nearest obstacle is the Euclidean distance to the single nearest map point found by an octree search, and the gradient is the finite difference of that distance along each axis. Because the field is generated on demand rather than maintained over the whole map, the system never performs obstacle inflation or a BFS-based ESDF update. The other mechanism is the adaptive high-speed trajectory initialization: waypoints are sampled uniformly along each topological path, and the per-segment durations are set so the drone accelerates to a desired speed $v_d$, flies at that speed, then decelerates, with $v_d$ iteratively scaled down until the path length can physically support the maneuver.","core_discovery":"The central claim is that high-speed obstacle avoidance does not require a global Euclidean signed distance field or a single trajectory optimization; it can be achieved by maintaining a robocentric point cloud that returns distance and gradient on demand (RESDF), enumerating multiple topologically distinct collision-free paths with a visibility-plane search, optimizing each path with an adaptive initialization, and selecting the trajectory with the smallest cost. Under this scheme the paper reports total per-iteration latencies of 4.5–7.2 ms in simulation and 90% success at a 15 m/s speed limit in an extremely dense forest, with the planned trajectory close to the global optimum in both the temporal and spatial domains.","pith_inferences":["Beyond the paper: the 'near-global optimum' claim depends on the $K=4$ angular sampling in the topological search, so in narrow-gap environments a gap thinner than the sampling resolution could be invisible to the optimizer; counting path diversity against a denser sampling or an obstacle skeleton would test this.","Beyond the paper: the single-neighbor distance estimate could be made conservative by tracking local point density or neighbor radii, which would let the same architecture claim a formal safety margin rather than relying on the tuned threshold $d_s$.","Beyond the paper: the cost-based trajectory selection ignores trajectory length, a limitation the paper itself acknowledges; adding length or energy to the selection criterion would likely shorten the reported 54.11 m average path length at 15 m/s."],"forward_implications":["The planning loop completes in 4.53–7.22 ms across the reported simulation scenarios, so the same onboard computer can react to newly observed obstacles within one lidar frame.","At a 15 m/s speed limit in a simulated forest with roughly triple obstacle density, the system keeps a 90% success rate (17 of 20 flights) while reaching an average maximum speed of 13.82 m/s.","Removing the global ESDF update means the mapping cost stays low as map resolution increases; the reported total mapping time is about 10.5% of the second-best method in the benchmark.","Real-world flights in a cluttered indoor room and a forest reach 8.0 m/s and 11.0 m/s respectively, indicating that the system transfers from simulation to hardware."],"supporting_citations":[{"why":"Supplies the dynamic octree structure used for adding, deleting, and searching map points in the incremental robocentric map.","marker":"[17]"},{"why":"Serves as the robocentric occupancy-grid-mapping baseline that HEPP compares against for map update time, ESDF time, and memory use.","marker":"[12]"},{"why":"Serves as the fixed-origin ESDF mapping baseline against which the on-demand RESDF approach is benchmarked.","marker":"[8]"},{"why":"Serves as the octree-based occupancy-grid baseline for mapping efficiency and memory comparison.","marker":"[14]"},{"why":"Supplies the PRM-based topological path search and trajectory replanning baseline used for path diversity and high-speed planning comparisons.","marker":"[5]"},{"why":"Supplies the A*-based topological path search baseline and the multi-path planning comparison.","marker":"[29]"},{"why":"Supplies the polynomial trajectory class with time-space deformation that the back-end trajectory optimization is built on.","marker":"[39]"},{"why":"Supplies the lidar-inertial odometry used to estimate the drone state and to transform scans into the body frame.","marker":"[46]"}],"fun_headline_variants":["Drones dodge obstacles at 15 m/s with a 79% faster planner","UAV replanning cut by 79% at 15 m/s in cluttered forests","Near-optimal quadrotor paths in milliseconds at 15 m/s","Hyper-efficient perception and planning boosts UAV speed to 15 m/s","89% faster mapping, 79% faster planning: drones fly fast and safe"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper treats the distance to the single nearest point in a sparse, sliding point cloud as the true distance to the nearest obstacle surface, and every collision check in the optimizer uses that number.","fun_headline_variants_meta":{"raw":{"variants":["Drones dodge obstacles at 15 m/s with a 79% faster planner","UAV replanning cut by 79% at 15 m/s in cluttered forests","Near-optimal quadrotor paths in milliseconds at 15 m/s","Hyper-efficient perception and planning boosts UAV speed to 15 m/s","89% faster mapping, 79% faster planning: drones fly fast and safe"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1405,"prompt_tokens":910,"completion_tokens":495,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":526,"completion_tokens_details":{"reasoning_tokens":390}},"tokens_in":526,"tokens_out":495,"duration_ms":3513,"temperature":1.0,"reasoning_tokens":390,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:47:00.717828+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline in a cluttered forest while computing a ground-truth signed distance field from dense registered scans; if any trajectory that the planner accepts has true clearance below the safety threshold $d_s$, or if the mean error between RESDF and the ground-truth distance at sampled trajectory points is comparable to $d_s$, the nearest-neighbor distance assumption is falsified.","supporting_citations":[{"cited_title":"i-octree: A fast, lightweight, and dynamic octree for proximity search,","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic octree structure used for adding, deleting, and searching map points in the incremental robocentric map."},{"cited_title":"Rog-map: An efficient robocentric occupancy grid map for large-scene and high- resolution lidar-based motion planning,","cited_arxiv_id":null,"evidence_quote":"Serves as the robocentric occupancy-grid-mapping baseline that HEPP compares against for map update time, ESDF time, and memory use."},{"cited_title":"Fiesta: Fast incremental euclidean distance fields for online motion planning of aerial robots,","cited_arxiv_id":null,"evidence_quote":"Serves as the fixed-origin ESDF mapping baseline against which the on-demand RESDF approach is benchmarked."},{"cited_title":"Raptor: Robust and perception- aware trajectory replanning for quadrotor fast flight,","cited_arxiv_id":null,"evidence_quote":"Supplies the PRM-based topological path search and trajectory replanning baseline used for path diversity and high-speed planning comparisons."},{"cited_title":"Ego-swarm: A fully autonomous and decentralized quadrotor swarm system in cluttered environments,","cited_arxiv_id":null,"evidence_quote":"Supplies the A*-based topological path search baseline and the multi-path planning comparison."},{"cited_title":"Fast-lio2: Fast direct lidar-inertial odometry,","cited_arxiv_id":null,"evidence_quote":"Supplies the lidar-inertial odometry used to estimate the drone state and to transform scans into the body frame."}],"review_version":1}