{"id":"2f3c227a-b953-4083-9902-ae51d9fad41a","arxiv_id":"2602.01189","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SPOT uses a 4D spatio-temporal RRT* tree, vision-based safe corridors, and a reactive backup planner to navigate a quadrotor around dynamic obstacles without a map.","lead":"SPOT is a quadrotor planner that combines time-aware search with vision-based safe corridors and a backup maneuver to avoid moving obstacles without building a map. It reports higher success rates than two prior planners in simulations with up to thirty moving obstacles, and demonstrates avoidance flights on a real drone.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I percentages incompatible with stated 50-trial protocol","rationale":"The reader's weakest_assumption focused on the constant-velocity prediction model (Eq. 5) and the treatment of untracked points as static. That is a legitimate generalization concern, but the paper's own text acknowledges future work on motion-model improvement, and the simulation experiments use constant-velocity obstacles, so the central simulation claim is internally consistent under that model. My stress-test found a more direct and concrete problem: the success rates in Table I are arithmetically incompatible with the stated '50 independent trials' protocol. This is an internal inconsistency that affects the strongest quantitative claim of the paper. It does not necessarily imply fabrication—the authors may have averaged across multiple runs or used a different number of trials than stated—but it means the empirical evidence as presented cannot be verified without additional raw data. Because this is fixable by providing trial counts or corrected statistics, the appropriate verdict remains CONDITIONAL, not outright rejection. I therefore leave the reader's verdict unchanged, while adding a specific condition: the raw trial counts must be supplied and shown consistent with the reported percentages. My agreement is partial because the reader correctly identified quantitative evidence as a weak point, but did not pinpoint the specific impossible-percentage issue.","tokens_in":10556,"tokens_out":8124,"duration_ms":94147,"concrete_test":"Request the raw per-trial success/failure logs (or exact integer counts) for each cell of Table I. For each reported percentage p, check whether p*50 is an integer. If p*50 is not an integer (e.g., 0.802*50=40.1), then the value cannot come from 50 independent binary trials; the authors must clarify the actual number of trials or the averaging procedure. If raw logs are unavailable, re-run the three obstacle-density configurations with 50 fixed random seeds per method, record the number of successes out of 50, and compare the resulting multiples of 2% to Table I. This settles whether the headline empirical claim is reproducible under the stated protocol.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim of the paper—that SPOT with backup planning achieves high success rates and outperforms baselines—rests entirely on Table I. However, the reported percentages cannot be derived from the stated protocol. Section IV.A.1 says 'For each environment, 50 independent trials are performed.' With 50 binary trials, any success rate must be a multiple of 2% (k/50). Table I reports values such as 80.2%, 94.2%, 71.3%, 57.1%, 62.2%, 52.2%, 42.8%, and 10.2%. None are multiples of 2%. For example, SPOT at 30 obstacles is 80.2%, which would be 40.1 successes out of 50; SPOT-w/o backup at 20 obstacles is 82.4%, which would be 41.2 out of 50. These are impossible for individual sets of 50 binary trials. The numbers might be averages over multiple batches or interpolated from some other procedure, but no such procedure is described. This is an internal inconsistency in the strongest quantitative evidence, not a matter of interpretation or benchmark choice. The 'more than 20%' improvement claim from backup planning is also derived from these numbers, so the central result is not reproducible from the stated experimental design.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SPOT, a mapless reactive planning framework for quadrotor navigation in unknown environments with dynamic obstacles. The pipeline uses a 4D spatio-temporal RRT* planner over point clouds, with dynamic points propagated by a constant-velocity model, followed by spatio-temporal Safe Flight Corridor construction and MINCO-based trajectory optimization. A backup module is added to handle deadlocks. Validation is performed in simulation against two external baselines and in hardware experiments. The central claim is that SPOT with backup planning achieves higher success rates than the cited baselines in dense dynamic environments (Table I), and that the backup module improves success rates by more than 20%.","tokens_in":10903,"tokens_out":4754,"duration_ms":51066,"significance":"If the empirical results are correct, the paper makes a useful contribution: a mapless, vision-based dynamic obstacle avoidance pipeline, a backup-planning mechanism for deadlock situations, and a public code release. The hardware experiments and external-baseline comparisons are also strengths. However, the central quantitative evidence is compromised by an internal inconsistency in the reported success rates, which are impossible under the stated 50-trial protocol. Because the contribution claims are largely empirical, the paper cannot be accepted in its current form; the issues are fixable by rerunning/reporting the experiments correctly, so major revision is appropriate.","major_comments":[{"comment":"The text states that 'For each environment, 50 independent trials are performed.' With binary success per trial, any success rate must be a multiple of 2% (k/50). Table I reports 80.2%, 94.2%, 71.3%, 57.1%, 62.2%, 52.2%, 42.8%, and 10.2%, none of which is a multiple of 2%. For example, SPOT at 30 obstacles would need 40.1 successes out of 50, and SPOT-w/o backup at 20 obstacles would need 41.2. This is not a statistical nuance; it means the central empirical result cannot be reproduced from the described experimental design. Please report raw counts or, if the numbers are averages over a different protocol, describe that protocol explicitly and provide confidence intervals or statistical tests.","section":"Section IV.A.1, Table I"},{"comment":"The claim that 'incorporating a backup trajectory increases the success rate by more than 20%' is only true for the 30-obstacle case (80.2 vs 52.2, a 28.0 percentage-point gain). At 10 and 20 obstacles the gains are 5.8 and 9.6 percentage points, respectively. Additionally, Table II reports that at 10 obstacles the backup is invoked 0 times on average, yet Table I attributes a 5.8-point difference to the backup module at that density; with zero invocations, the difference is unexplained. This inconsistency further burdens the backup-contribution claim.","section":"Section IV.A.1, text after Table I"},{"comment":"The baseline comparison is not fully matched: SPOT and [14] run at 1 m/s, while [7] is reported at both 1 m/s and 5 m/s. Since [7] is designed for high-speed navigation, the headline comparison should be against [7]-1m/s, and the [7]-5m/s row should not be used to argue that SPOT outperforms [7]. Also, no confidence intervals, standard deviations, or significance tests are provided for the 50-trial success rates. The current Table I does not support a claim of statistically significant superiority over the baselines.","section":"Section IV.A.1, comparison protocol"},{"comment":"The safety claims are conditional on assumptions that are not quantified: (i) every point not inside a tracked bounding box is classified as static, (ii) unexplored regions outside the sensor FOV are assumed obstacle-free, and (iii) obstacle motion follows the constant-velocity model of Eq. (5). The paper presents the achieved trajectories as 'collision-free' without reporting tracker recall, detection failure rates, or prediction-error statistics in the pipeline validation. These assumptions should be stated as limitations, and, if possible, the PX4 SITL validation should report how often the onboard detector missed or mis-tracked an obstacle.","section":"Section III-A and III-C"}],"minor_comments":[{"comment":"The term 'success rate' is not formally defined. Is it 'reaching the goal without collision' or something else? Please state the criterion explicitly.","section":"Section IV.A.1"},{"comment":"The hardware results are qualitative (Fig. 9). To substantiate the claim of 'reliable performance using only onboard sensing and computation,' please include quantitative metrics such as trajectory tracking error, minimum distances to obstacles, or success/failure counts.","section":"Section IV.B"},{"comment":"Typos: 'All simulations were were executed' (Section IV.A), 'reachible' (Section III.B.2), 'MeitY ,' with an extra space (Acknowledgments). Please proofread.","section":"Throughout"},{"comment":"The informed sampling in Eq. (3) is defined in 3D space, and the time coordinate is added via Eq. (4). This is a decoupled heuristic rather than a true 4D informed set. Please clarify this in the text to avoid overstating the extension.","section":"Section III.B.3"}],"recommendation":"major_revision","confidential_remarks":"The paper has a plausible architecture and useful engineering contributions, but the Table I inconsistency is a red flag that must be resolved before the empirical claims can be trusted. I recommend major revision, not rejection, because the issue appears fixable by rerunning the experiments and reporting the data correctly. Please also require the authors to address the baseline-speed mismatch and the unsupported >20% backup-improvement statement."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know first: the paper is a coherent system paper, not a breakthrough. The new bit is the integration — time-stamped RRT* nodes with mapless point-cloud safe flight corridors, plus a deadlock-triggered backup planner. That backup mode is a real practical contribution; the ablation against 'SPOT-w/o backup' shows it matters. The benchmark comparisons are against external methods, and the hardware results, while qualitative, at least show the pipeline runs on a real quadrotor. Credit where due: the planning chain is sensible, the constant-velocity prediction is standard for short horizons, and the writing is clear about what is and isn't assumed.\n\nThe problem is Table I. The text says 50 independent trials per environment. With 50 binary trials, success rates must be multiples of 2% (k/50). The table reports 80.2%, 94.2%, 71.3%, 57.1%, 62.2%, 52.2%, 42.8%, and 10.2% — none of which are multiples of 2. For example, 80.2% would be 40.1 successful trials out of 50, which is impossible for a single set. This isn't a minor typo; the central claim — that SPOT outperforms baselines and that backup improves success by over 20% — rests entirely on these numbers. The authors may have averaged across batches or used a different trial count, but they don't say so, and we can't reproduce the result from the stated protocol.\n\nThat's the load-bearing flaw. To a lesser degree: no confidence intervals or significance tests around the percentages, one baseline ([7]) is run at 1 and 5 m/s while others run at 1 m/s, making that comparison uneven, and the hardware section is illustrative rather than measured. None of these are fatal on their own, and the algorithmic idea is sound.\n\nThe honest summary: this is a useful system with a well-motivated backup mechanism, but the empirical evidence as presented is not trustworthy. The fix is straightforward — release raw trial counts, correct the table, add error bars — but it has to happen before the numbers are used.\n\nMy recommendation: yes, send it to peer review. The system deserves referee time, but the authors need to either provide the raw data or revise the experimental reporting. I'd bring this to a reading group to discuss the backup mechanism, and I'd cite the algorithmic idea, not the reported success rates.","headline":"Nice integration of known pieces with a genuinely useful deadlock fallback, but Table I's success rates can't be derived from 50-trial runs — that needs fixing before we can trust the headline numbers.","tokens_in":11352,"tokens_out":1772,"would_cite":false,"duration_ms":23040,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T40"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a mapless, time-aware path planner with backup trajectories lets quadrotors navigate unknown dynamic environments and raises success rates by over 20 percent in dense obstacle fields.","keywords":["quadrotor motion planning","dynamic obstacle avoidance","spatio-temporal RRT*","safe flight corridor","backup trajectory","mapless navigation","vision-based perception","deadlock recovery"],"falsifier":"Re-run the 30-obstacle benchmark with the backup module disabled and compare to the reported 80.2% success rate with backup on; the paper's claim of more than a 20-point gain fails if the gap does not reproduce. A second check: insert an obstacle that accelerates sharply within the 2-second prediction window and see whether the planner still avoids it.","tokens_in":1314,"feed_emoji":"🚁","tokens_out":4919,"duration_ms":99828,"temperature":0.7,"pith_summary":"The paper's core claim is that a quadrotor can navigate unknown, dynamic environments without a map by treating time as a fourth planning dimension and by adding a backup trajectory for deadlock situations. The proposed pipeline, SPOT, uses onboard depth perception to classify static versus dynamic points, builds a safe flight corridor from a spatio-temporal RRT* path, and optimizes a smooth trajectory inside it. In simulations with 10, 20, and 30 moving obstacles, the paper reports success rates of 100%, 92%, and 80.2%, and its ablation shows that the backup module alone raises the 30-obstacle success rate from 52.2% to 80.2%. If correct, this matters because it removes the need for map fusion or motion-capture infrastructure, making reactive UAV navigation feasible with onboard sensing and computation alone.","feed_headline":"Backup trajectories boost drone success rate by over 20 percent","feed_subtitle":"A mapless 4D planner lets UAVs dodge moving obstacles and escape deadlocks using only onboard sensing.","key_machinery":"The load-bearing object is the 4D node state n_i=(x,y,z,t_i) in a spatio-temporal RRT*: arrival time is propagated from parent to child, and the feasibility set F(n_i) requires the child's arrival time to be strictly later than the parent's and reachable under a maximum velocity bound. Collision checking queries a spatio-temporal hash grid that stores where each dynamic point will be at discrete times under constant-velocity prediction, together with a kd-tree for static points. The safe flight corridor is constructed by aggregating static point clouds with dynamic point clouds predicted over the interval between nodes, then performing convex decomposition; trajectory optimization minimizes","core_discovery":"In the paper's own terms, SPOT establishes that reactive planning in unknown dynamic environments can be done by augmenting RRT* with arrival time at every node. Each node n_i=(x,y,z,t_i) is checked for collision at its predicted arrival time against a spatio-temporal hash grid built from detected moving obstacles, while static points form a kd-tree. The resulting path is inflated into overlapping convex polyhedra over the time interval, with predicted obstacle point clouds from a constant-velocity model aggregated into the obstacle set. Trajectory optimization then fits a minimum-jerk polynomial inside the corridor. When no goal-directed path is available, a backup module picks an escape di","pith_inferences":["My inference: the 2-second constant-velocity prediction horizon sets an upper bound on obstacle acceleration; in scenes with sudden stops or darting objects, the planner would likely need learned motion models or uncertainty-aware prediction.","My inference: a natural stress test is occlusion—an obstacle hidden behind another until it enters the field of view; the paper's assume-free-for-unexplored-space rule means such an object is invisible until detected, so performance likely depends heavily on tracker recall.","My inference: the backup-escape mechanism transfers beyond UAVs—any mobile robot with a convex free-space certificate and an escape direction could use the same deadlock-avoidance loop; a direct test would be deploying it on a ground robot with 2D lidar."],"forward_implications":["With only onboard depth sensing and odometry, a quadrotor can avoid 10 to 30 moving obstacles without a fused map or external motion capture.","Deadlock situations become survivable: instead of hovering in place, the UAV executes an obstacle-repelling backup trajectory and re-enters goal-directed planning once a path exists.","The success-rate gain concentrates in dense scenes: at 30 obstacles, reported success is 80.2% with backup versus 52.2% without, so backup planning is what makes dense dynamic navigation practical.","Because the spatio-temporal feasibility set naturally extends to bounded acceleration, the same planner can be adapted to different vehicle agility limits.","Since the safe flight corridor is built only within the sensing horizon and unexplored space is assumed free, computation stays bounded enough for onboard replanning."],"fun_headline_variants":["Mapless UAV planner uses arrival time to avoid moving threats","SPOT planner escapes deadlocks with reactive backup path","4D spatio-temporal planning dodges unknown dynamic obstacles","Vision-only UAV avoids collisions via backup escape module","Reactive 4D planning with backup paths boosts UAV success"],"cache_read_input_tokens":12800,"weakest_assumption_plain":"The load-bearing premise is that every point outside a detected bounding box is static and every tracked obstacle moves at constant velocity for up to two seconds; if the detector misses an object or an obstacle accelerates inside that horizon, the planner's occupancy model is wrong and the nominal safe trajectory can collide.","fun_headline_variants_meta":{"raw":{"variants":["Mapless UAV planner uses arrival time to avoid moving threats","SPOT planner escapes deadlocks with reactive backup path","4D spatio-temporal planning dodges unknown dynamic obstacles","Vision-only UAV avoids collisions via backup escape module","Reactive 4D planning with backup paths boosts UAV success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000433,"raw_usage":{"total_tokens":2012,"prompt_tokens":683,"completion_tokens":1329,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":427,"completion_tokens_details":{"reasoning_tokens":1263}},"tokens_in":427,"tokens_out":1329,"duration_ms":11232,"temperature":1.0,"reasoning_tokens":1263,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T05:44:07.931356+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the 30-obstacle benchmark with the backup module disabled and compare to the reported 80.2% success rate with backup on; the paper's claim of more than a 20-point gain fails if the gap does not reproduce. A second check: insert an obstacle that accelerates sharply within the 2-second prediction window and see whether the planner still avoids it.","supporting_citations":[],"review_version":1}