{"id":"ed6e555a-eeb2-4a0f-99d7-d02d5a5ab815","arxiv_id":"2504.17569","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A LiDAR-based UAV system that detects moving points, predicts obstacle trajectories, and integrates those predictions into a 100 Hz MPC planner, achieving 0.97-1.00 simulation success rates and real-world forest flights.","lead":"This paper presents a complete LiDAR-based system that lets a drone detect, predict, and avoid moving obstacles while flying through cluttered environments. It combines the authors' prior M-detector perception and IPC planning into a new DynIPC controller, and demonstrates flights through forests with thrown balls, pendulums, cars, and pedestrians.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on an unvalidated polynomial obstacle-motion model: the benchmark obstacles are reciprocating, which neither constant-velocity nor free-fall prediction represents, and no ablation or prediction-error statistics show that DynIPC rather than reactive replanning…","rationale":"The paper's central claim is a complete LiDAR-based onboard stack that avoids fast-moving, suddenly appearing, heterogeneous obstacles, with high simulation success rates and 11.54 ms real-world latency. For that claim to be true, the dynamic-object prediction module does not have to be perfect, but the paper's novelty (DynIPC over IPC) is specifically that predictions enable avoidance 'in advance' (Sec. II-B, VI). The benchmark scenarios include reciprocating motion, and the pendulum experiment, both of which are poorly matched to the constant-velocity/free-fall family. The paper reports no measuring of prediction error, no case where the model failed, and no ablation of the prediction module. This is a load-bearing gap: without it, the high success rates may be entirely due to the reactive 100 Hz IPC loop, which already existed. The reader's weakest assumption identifies the same issue. The proposed ablation test would settle it. I do not see an internal inconsistency that would overturn the system-level demonstration, so I recommend keeping the CONDITIONAL verdict and making the DynIPC ablation and prediction-error reporting explicit conditions.","tokens_in":12192,"tokens_out":8908,"duration_ms":96923,"concrete_test":"Run Scene 1 with the same 30 runs but with the prediction module ablated: use only instantaneous obstacle positions in the dynamic layer and Eq. (2) instead of the 2 s polynomial predictions, and compare success rate, minimum clearance, and replanning frequency to Table I. If the 0.97 success rate is unchanged, the polynomial model is not load-bearing; if it degrades materially, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest load-bearing link is the trajectory-prediction assumption in Sec. IV-B. The system fits each tracked obstacle's future motion over a 2 s window with only two models, constant velocity (linear fit) or free fall (quadratic fit with coefficient near g/2), and both the dynamic occupancy layer (Sec. V) and the avoidance cost in Eq. (2) are built from those predicted positions. The simulation obstacles in Sec. VII-A are stated to be in 'reciprocating motion at 6 m/s' (Scene 1) and 2 m/s (Scene 2), and the real-world demonstration includes a pendulum. A reciprocating or pendulum trajectory reverses direction on a timescale shorter than the prediction/MPC horizon, so a sliding-window polynomial fit will systematically place the predicted obstacle on the wrong side of the reversal. The paper does not report prediction-error statistics, does not compare DynIPC against its own reactive IPC baseline, and does not ablate the prediction module. Consequently the reported success rates (0.97 and 1.00) and the 11.54 ms latency do not establish that predictive avoidance is what makes the system work; the 100 Hz reactive replanning loop acknowledged in Sec. VI-D may be doing the safety work. This does not falsify the system-level demonstration, but it leaves the central DynIPC contribution unvalidated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a complete LiDAR-based UAV navigation system for cluttered and dynamic environments. The perception module uses the authors' prior M-detector to classify moving points, clusters and tracks dynamic objects with DBSCAN and the Hungarian algorithm, and predicts their future positions over a 2 s horizon using either constant-velocity or free-fall polynomial fits. The planning module extends the authors' prior IPC framework into DynIPC by adding dynamic-obstacle predictions to a two-layer ROG-Map (static and dynamic layers) and to an MPC-based integrated planning-and-control problem, solved at 100 Hz. The system is evaluated in two Gazebo simulation scenes against Panther and Chen's work, where it reports higher success rates (0.97 and 1.00), lower flight times, and higher maximum velocities, and in a real-world forest experiment with a LIVOX Mid-360 LiDAR, where it reports a perception-to-control latency of 11.54 ms and successful avoidance of cars, balls, pedestrians, and pendulums.","tokens_in":12517,"tokens_out":5061,"duration_ms":49479,"significance":"If the claims hold, this would be a useful system-level demonstration: a single LiDAR-based onboard stack that detects and avoids heterogeneous moving obstacles in real time. The paper's strengths include the integration of previously developed components into a complete pipeline, the use of two open-source baselines for comparison, and a real-world test with diverse dynamic obstacles. The low reported latency and the high simulation success rates indicate practical potential. However, the evidence as presented does not yet establish the central contribution of predictive avoidance: the simulation substitutes a depth camera for the LiDAR sensor, the trajectory prediction model is not validated against the actual obstacle motions used in the tests, and the real-world validation is a single qualitative flight. The paper is clearly written and the supplementary video is a positive asset.","major_comments":[{"comment":"The simulation experiments replace the LiDAR sensor with a depth camera for all methods, including the proposed system, and use ground-truth localization. This means the simulation results do not exercise the LiDAR-based perception module (M-detector), the LiDAR odometry, or the sensor fusion pipeline that the paper's title and abstract claim to validate. The real-world experiment in Section VII-B does use the LIVOX Mid-360, but it is a single qualitative demonstration without repeated trials or quantitative success statistics. To support the central claim of a LiDAR-based system, the authors should either run the simulation with a LiDAR sensor model (even at the cost of benchmark comparability) or explicitly reposition the real-world experiment as the primary validation of the LiDAR pipeline and provide repeated-trial statistics.","section":"VII-A"},{"comment":"The trajectory prediction module assumes that each tracked obstacle follows either constant-velocity or free-fall motion over a 2 s window, yet the simulation obstacles are described as being in reciprocating motion at 6 m/s and 2 m/s, and the real-world experiment includes a pendulum. A reciprocating or pendulum trajectory reverses direction within the prediction/MPC horizon, so the fitted linear or quadratic curve will place the predicted obstacle on the wrong side of the reversal. Both the dynamic occupancy layer (Section V) and the avoidance cost in Eq. (2) are built from these predicted positions. The paper reports no prediction-error statistics, no ablation comparing DynIPC with the reactive IPC baseline, and no analysis of cases where the prediction assumption failed. These omissions leave the central predictive-avoidance contribution unvalidated, because the 100 Hz reactive replanning loop described in Section VI-D could itself account for the reported success rates.","section":"IV-B, VII-A"},{"comment":"The simulation results in Table I report point estimates over 30 runs without variance, confidence intervals, or statistical tests, and the experimental setup does not state what randomness, if any, differentiates the runs. With success rates of 0.60 versus 0.97 and 0.07 versus 1.00, the claim that the system outperforms the baselines needs at least standard deviations or a description of the stochastic variations. In addition, the real-world section reports a single flight; the 11.54 ms latency and the 4.55 m/s maximum velocity are single measurements with no repeated trials. The empirical evidence is suggestive but not yet convincing at the level of a journal publication.","section":"Table I, VII-B"}],"minor_comments":[{"comment":"The simplification from Eq. (1a) to Eq. (2) pre-selects the obstacles based on the reference trajectory position pref,n rather than the planned position pn. This should be stated more explicitly, because if the planned trajectory deviates from the reference path, the set of active obstacle costs can be wrong; the current wording in the text describes the reference position as \"the estimated optimal UAV position\" without justifying that approximation.","section":"VI-C, Eq. (2)"},{"comment":"The paper says M-detector is employed in the simulation, but M-detector was originally developed and validated for LiDAR point streams. The authors should clarify whether M-detector requires any parameter changes or recalibration when applied to depth-camera point clouds, since the simulation uses a depth camera for all methods.","section":"VII-A"},{"comment":"Table II lists vi,max = 4 as the maximum velocity in the x, y, z direction, but Table I reports maximum velocities of 4.87 m/s and 4.56 m/s. If vmax in Table I is the Euclidean norm of the velocity, this is consistent, but the distinction between per-axis limits and the reported norm should be stated explicitly to avoid confusion.","section":"Table II"},{"comment":"The term \"integrated planning and control\" is used both for the overall system and for the specific MPC module (DynIPC). The paper would benefit from a consistent terminology, e.g., reserving \"integrated planning and control\" for the MPC-based back end and using \"system\" for the complete pipeline.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' own prior work (M-detector, IPC, ROG-Map), which is acceptable, but the novelty of DynIPC relative to IPC should be more sharply delineated. The sensor mismatch in simulation (depth camera instead of LiDAR) is likely to provoke strong reviewer objections and should be addressed head-on with additional experiments or a clear re-framing of the claims. The lack of prediction-error evaluation is the deepest technical gap and should be prioritized in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: this is a competent systems paper, and the real-world flights are credible, but the headline claim that DynIPC's predictive avoidance is what makes it work is not supported by the evidence. The evaluation never isolates the contribution.\n\nThe genuinely new bit is DynIPC—extending the authors' earlier IPC planner with a soft cost for predicted positions of dynamic obstacles, plus static-only SFC and a dynamic-aware A* search. Putting M-detector, the two-layer ROG-Map, and DynIPC together into one onboard LiDAR stack is a useful engineering contribution. The real data are meaningful: 11.54 ms perception-to-control latency, flights through a forest with pendulums, balls, a car, and pedestrians. That is not nothing.\n\nThe weak link is exactly the trajectory-prediction module. Section IV-B fits only constant-velocity or free-fall polynomials over a 2 s window. The simulation obstacles are in reciprocating motion (6 and 2 m/s); a reciprocating trajectory reverses direction faster than the prediction horizon, so a sliding-window polynomial fit is systematically wrong around the turn. Yet the paper never reports prediction error, never ablate the prediction module, and never compares DynIPC against its own reactive IPC baseline with the same stack. The success rates (0.97 and 1.00) and the latency figure therefore do not establish that predictive avoidance is doing the safety work; the 100 Hz reactive replanning loop could be. This is a load-bearing gap in the paper's scientific claim, though not in the system demonstration.\n\nOther soft spots, mostly minor: the simulation uses a depth camera instead of LiDAR for all methods, so the \"LiDAR-based system\" claim is only directly tested in the single real-flight sequence. Table I reports no variance over the 30 runs. The convexity assertion—setting Ro < Rp gives a positive semidefinite Hessian—is stated without proof; it may be true but it is not obvious from Eq. (2) as written. The citation pattern is heavily self-referential, but for an integration paper that is appropriate; the prior work is published and appears real.\n\nBottom line: this is a worthwhile systems paper for the UAV navigation community, and it deserves referee time. But the referee should ask for an ablation with the dynamic cost removed and for prediction-error statistics before the DynIPC contribution is accepted. If those come back, the paper becomes a solid reference architecture.","headline":"A credible systems integration whose key new component (DynIPC) is never isolated in the evaluation—the predictive-avoidance claim needs an ablation and prediction-error statistics before it lands.","tokens_in":13060,"tokens_out":3358,"would_cite":false,"duration_ms":32737,"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 LiDAR-only system can fly a drone through cluttered forests while avoiding fast-moving obstacles onboard.","keywords":["LiDAR-based UAV navigation","dynamic obstacle avoidance","integrated planning and control","moving object detection","trajectory prediction","occupancy grid map","cluttered and dynamic environments","real-time onboard flight"],"falsifier":"A controlled flight test with an obstacle that accelerates or turns sharply inside the two-second prediction horizon, in otherwise identical forest conditions, would settle the matter: if near-miss distances stay the same, the reactive replanner is doing the safety work; if collisions or unsafe clearances appear, the polynomial motion assumption is the load-bearing premise.","tokens_in":11996,"feed_emoji":"🚁","tokens_out":7042,"duration_ms":63089,"temperature":0.7,"pith_summary":"This paper makes the case that a single LiDAR sensor, an onboard computer, and the right division of labor between perception and planning are enough for a drone to fly through cluttered, dynamic surroundings. The system detects moving points without relying on object classes or colors, predicts where each tracked obstacle will be over a two-second horizon, and folds those predictions into a 100 Hz integrated planner-controller so it can dodge fast or suddenly appearing objects. In simulation it reaches success rates of 0.97 and 1.00 in two scenes, beating two open-source baselines, and in a forest it avoided thrown balls, pendulums, cars, and pedestrians with a measured perception-to-control latency of 11.54 ms. If the claims hold, LiDAR-only drones can operate safely in environments that previously required cameras, trained detectors, or reactive-only avoidance.","feed_headline":"LiDAR-only drone dodges moving obstacles in 11.5 ms","feed_subtitle":"Onboard perception predicts where obstacles will be, letting a drone avoid fast-moving objects in a forest.","key_machinery":"The load-bearing mechanism is DynIPC, an integrated planning-and-control formulation in which the predicted position of each tracked obstacle at every MPC step enters the cost as $\\max(d_{\\mathrm{dyn}}^2 - \\|p_n - p_{\\mathrm{obs},n,k}\\|^2_{R_o}, 0)$, simplified to a negative quadratic so the whole problem stays a convex QP solvable at 100 Hz. Around it sit two supporting mechanisms: a two-layer occupancy map (static points only for the static layer, predicted obstacle footprints for the dynamic layer, with velocity-scaled inflation $d_{\\mathrm{inf}} = d_{\\mathrm{rad}} + k\\cdot\\mathrm{vel}$) and the M-detector frame-out pipeline that supplies the moving points and trajectory fits. Together they let the planner treat static clutter as hard constraints and dynamic obstacles as soft, smoothly avoiding many moving objects without resorting to nonlinear optimization.","core_discovery":"The central claim is that dynamic obstacle avoidance can be solved onboard with LiDAR alone by separating the world into a static occupancy layer built only from static points and a dynamic layer built from predicted future positions of moving objects. Moving points are extracted by M-detector's frame-out mode, clustered with DBSCAN, associated across frames with the Hungarian algorithm, and fit to one of two motion models—constant velocity or free fall—to forecast positions up to two seconds ahead. The planner, DynIPC, extends the authors' earlier integrated planning and control work so that these predicted positions enter a convex MPC as soft quadratic penalties, while static obstacles remain hard corridor constraints; the same algorithm also searches the reference path with A* over grids that are both statically and dynamically safe. The paper's own summary of the result is that with low-latency perception and prediction-aware planning, the UAV avoids heterogeneous moving obstacles in cluttered real environments while all computation runs onboard.","pith_inferences":["The two-motion-model prediction (constant velocity or free fall) is the part most likely to limit real generality; a drone, car, or animal that turns or brakes within the two-second horizon would invalidate the predicted positions, and the system would then depend on the reactive replanning loop for safety.","The same static/dynamic map split and soft-constraint MPC could be transferred to ground robots or manipulation with point-cloud sensing, since nothing in the formulation is quadrotor-specific except the kinodynamic limits.","One testable extension is to replace the fixed velocity-scaled inflation with uncertainty-aware inflation derived from prediction covariance, which would make the safety margin adapt to how confident the tracker is.","The simulation comparison used a depth camera in all three systems to isolate the planning contribution; real-world results use LiDAR, so an end-to-end LiDAR simulation benchmark with the same baselines would directly test the sensor advantage."],"forward_implications":["A LiDAR-only sensing suite, without cameras or pretrained object classes, is sufficient for cluttered dynamic navigation outdoors.","Fast, sudden-appearing obstacles can be handled at up to 8.75 m/s relative speed because the full perception-to-control chain runs in about 11 ms.","Because dynamic avoidance is a soft quadratic cost in a convex QP, many tracked obstacles can be considered simultaneously without breaking the 100 Hz control rate.","Using only static points to build safe flight corridors enlarges the navigable space and prevents ghost points from moving objects from blocking the path.","In the paper's comparisons, prediction-aware planning reaches the goal faster and at higher speed than reactive or learning-based baselines while maintaining a higher success rate."],"supporting_citations":[{"why":"Supplies the M-detector moving-point classification that separates static and dynamic points at low latency.","marker":"[13]"},{"why":"Provides the integrated planning and control base, including SFC generation and MPC formulation, that DynIPC extends.","marker":"[14]"},{"why":"ROG-Map is the occupancy map structure extended into static and dynamic layers for the system.","marker":"[19]"},{"why":"FAST-LIO2 provides the 100 Hz LiDAR-inertial odometry used for localization in real-world experiments.","marker":"[22]"},{"why":"DBSCAN clusters moving points into object clusters before tracking and prediction.","marker":"[17]"},{"why":"The Hungarian algorithm performs frame-to-frame association of multiple dynamic objects.","marker":"[18]"},{"why":"A* searches the reference path over grids that are both statically and dynamically safe.","marker":"[20]"},{"why":"OSQP solves the convex quadratic program at the core of the MPC.","marker":"[21]"},{"why":"Panther is a baseline system compared in simulation, providing the depth-camera dynamic environment avoidance reference.","marker":"[5]"},{"why":"Chen's work is the learning-based baseline compared in simulation, providing ground-truth perception for the benchmark.","marker":"[7]"}],"fun_headline_variants":["LiDAR drone dodges moving obstacles in real time","Onboard LiDAR predicts and avoids moving obstacles","UAV flies through forests dodging dynamic obstacles","Real-time LiDAR system for cluttered dynamic skies","Drone with LiDAR avoids fast-moving obstacles online"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Predictive avoidance rests on the assumption that every tracked obstacle keeps moving with constant velocity or in free fall for the next two seconds; an obstacle that turns, brakes, or is pushed will not be where the map and cost function expect it to be.","fun_headline_variants_meta":{"raw":{"variants":["LiDAR drone dodges moving obstacles in real time","Onboard LiDAR predicts and avoids moving obstacles","UAV flies through forests dodging dynamic obstacles","Real-time LiDAR system for cluttered dynamic skies","Drone with LiDAR avoids fast-moving obstacles online"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000183,"raw_usage":{"total_tokens":1308,"prompt_tokens":929,"completion_tokens":379,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":545,"completion_tokens_details":{"reasoning_tokens":305}},"tokens_in":545,"tokens_out":379,"duration_ms":3559,"temperature":1.0,"reasoning_tokens":305,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:36:24.501446+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled flight test with an obstacle that accelerates or turns sharply inside the two-second prediction horizon, in otherwise identical forest conditions, would settle the matter: if near-miss distances stay the same, the reactive replanner is doing the safety work; if collisions or unsafe clearances appear, the polynomial motion assumption is the load-bearing premise.","supporting_citations":[{"cited_title":"Moving event detection from lidar point streams,","cited_arxiv_id":null,"evidence_quote":"Supplies the M-detector moving-point classification that separates static and dynamic points at low latency."},{"cited_title":"Integrated planning and control for quadrotor navigation in presence of suddenly appearing objects and disturbances,","cited_arxiv_id":null,"evidence_quote":"Provides the integrated planning and control base, including SFC generation and MPC formulation, that DynIPC extends."},{"cited_title":"Fast-lio2: Fast direct lidar- inertial odometry,","cited_arxiv_id":null,"evidence_quote":"FAST-LIO2 provides the 100 Hz LiDAR-inertial odometry used for localization in real-world experiments."},{"cited_title":"A density-based algorithm for discovering clusters in large spatial databases with noise,","cited_arxiv_id":null,"evidence_quote":"DBSCAN clusters moving points into object clusters before tracking and prediction."},{"cited_title":"The hungarian method for the assignment problem,","cited_arxiv_id":null,"evidence_quote":"The Hungarian algorithm performs frame-to-frame association of multiple dynamic objects."},{"cited_title":"A formal basis for the heuristic determination of minimum cost paths,","cited_arxiv_id":null,"evidence_quote":"A* searches the reference path over grids that are both statically and dynamically safe."},{"cited_title":"Flying in dynamic scenes with multitarget velocimetry and perception-enhanced planning,","cited_arxiv_id":null,"evidence_quote":"Chen's work is the learning-based baseline compared in simulation, providing ground-truth perception for the benchmark."}],"review_version":1}