{"id":"4a525423-23cc-48af-9757-29c42cc687b0","arxiv_id":"1908.02028","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A drone trajectory planner generates collision-free, near-time-optimal avoidance paths around moving rectangular obstacles in under 6 ms by optimizing waypoint velocities analytically.","lead":"Researchers at the University of Bonn present a planner that lets a drone dodge a moving box in a few milliseconds by inserting special waypoints into a time-optimal trajectory. The method works in simulation; its 'time-optimal' label rests on simplifying assumptions, and no code or hardware data is released.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'time-optimal' claim is only optimal within a restricted family: zero acceleration at obstacle-corner viastates, and the paper explicitly concedes nonzero acceleration can be faster.","rationale":"The reader's weakest_assumption already identifies the same load-bearing issue: time-optimality is claimed for a restricted family of trajectories through obstacle corners with zero acceleration, and the paper itself acknowledges nonzero acceleration can be faster. My stress-test agrees and adds that the Sec. IV-B.3 quadratic-polynomial intersection is an approximation rather than an exact solve, so the optimality claim is doubly weakened. This does not change the appropriate verdict: the paper remains a plausible algorithmic contribution with a useful reactive planner, but its headline optimality claim should be either proven for the full problem class, proven within the stated restricted family with exact optimization, or explicitly softened. The existing CONDITIONAL verdict already captures this; my read reinforces the conditions rather than altering them. No other concern seemed more load-bearing: the runtime claim (under 6 ms) is directly measured and plausible, the moving-obstacle handling is demonstrated qualitatively, and the method's lack of uncertainty handling is acknowledged and mitigated by replanning. The most important missing evidence remains a comparison against a global optimum or, failing that, a clear statement that 'time-optimal' means 'optimal within the zero-acceleration corner-viastate family'.","tokens_in":10969,"tokens_out":2909,"duration_ms":37272,"concrete_test":"For the scenario in Fig. 1 (and ideally all scenarios in Sec. V-B), compute the proposed trajectory and then solve a reference optimization with the same triple-integrator model and jerk/acceleration/velocity limits, allowing the viastate position to vary continuously on the obstacle boundary and the viastate acceleration to take any value in [-amax, amax]. Implement the reference via dense sampling of viastate velocity components plus nonlinear local refinement, or direct transcription with a small nonlinear program. Record the ratio of proposed total time to best-found total time. If any scenario has a ratio above 1.005 (i.e., a strictly faster feasible trajectory exists), the global time-optimality claim is refuted; if no scenario shows such a gap, the restriction is empirically supported for these cases, though still not proven.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the method 'maintain[s] time-optimality' (Abstract, Sec. IV-B.1) rests on the unproven assumption that the globally fastest collision-free trajectory in the considered problem class always passes through one of the four obstacle corners with zero acceleration at the viastate. The paper itself states in Sec. IV-B.1: 'a nonzero acceleration could be slightly faster ... but under the assumption of zero acceleration, the generated evading trajectories are optimal.' That is an explicit admission that the search family is restricted. No argument is given that the true optimum must lie at a corner, especially for moving cuboids with asymmetric start and target states. The multi-velocity tradeoff in Sec. IV-B.3 is also solved only approximately, by fitting second-order polynomials through three sample points (Eqs. 8-10), so even within the corner/zero-acceleration family the result is not proven to be the exact optimum. The evaluation (Sec. V-B) reports over 100 scenarios but provides no comparison against a global optimum, no optimality gap, and no statistics; the fuzz test of 10,000,000 trajectories covers partially defined target states (Sec. V-A), not obstacle avoidance. Consequently, the paper demonstrates 'optimal within a deliberately restricted family', while the abstract and conclusion make an unqualified global time-optimality claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a reactive obstacle-avoidance method for micro aerial vehicles (MAVs) built on the authors' earlier trajectory generator for third-order systems. The method first generates a time-optimal trajectory from a full start state to a possibly partially defined target state, checks it for collisions by exploiting the piecewise polynomial form, and, if a collision is found, inserts viastates at obstacle corners. For the two 'bound' axes, the viastate is placed at a projected obstacle corner with zero acceleration, and the viastate velocity is chosen by one of two heuristics: a single dominating velocity component, or a tradeoff computed from the intersection of quadratic polynomials fitted through three sample points. The remaining 'free' axes are synchronized to the bound axes. The paper reports sub-6 ms computation on a single CPU core, a fuzz test of partially defined target states, and over 100 obstacle-avoidance scenarios with one moving-obstacle example.","tokens_in":11300,"tokens_out":4791,"duration_ms":53994,"significance":"If the optimality claim held, the method would be a useful building block for high-speed MAV navigation: it is fast, handles moving axis-aligned cuboids, and produces smooth dynamically feasible trajectories with arbitrary dimensionality. The paper has concrete strengths: the collision check via zero crossings of polynomials is elegant, the solution database for partially defined target states is a practical contribution, the code is announced as open source, and the runtime claim is falsifiable and clearly stated. However, the headline contribution is currently the time-optimality guarantee, and that guarantee is not established: the search space is restricted to corner viastates with zero acceleration, and the multi-velocity tradeoff is solved by a polynomial fit rather than an exact optimization. The significance is therefore conditional on either restricting the claims to the considered family or providing a quantitative bound on the suboptimality gap.","major_comments":[{"comment":"The unqualified 'time-optimal' claim is not established. The search space for the viastate is restricted to obstacle corners with zero acceleration, and the paper itself states that 'a nonzero acceleration could be slightly faster' and that optimality holds only 'under the assumption of zero acceleration.' No argument or proof is given that the globally fastest trajectory in the considered problem class must pass through a corner, nor that zero acceleration is optimal; this is particularly non-obvious for moving cuboids with asymmetric start and target states. Since the Abstract and Conclusion repeat the unconditioned claim, the central contribution is currently over-stated. The authors should either restrict all optimality claims to the restricted family, or prove that the global optimum lies in that family and quantify any gap.","section":"Sec. IV-B.1, Abstract, Conclusion"},{"comment":"Even within the restricted corner/zero-acceleration family, the optimal viastate velocity is not computed exactly in the multi-velocity case. The method fits second-order polynomials through three sample points and solves for their intersection; the fitted polynomials are an approximation of the true cost landscape, and no error bound or validation against the sampled map (Fig. 4) is provided. The claim that the method 'maintains time-optimality' therefore depends on an unquantified approximation. Please either provide an exact derivation for these cases or evaluate the resulting suboptimality against dense sampling over a range of start/target/obstacle configurations.","section":"Sec. IV-B.3, Eqs. (8)-(10)"},{"comment":"The evaluation as reported is too weak to support the runtime and optimality claims. The 100+ scenarios are not summarized statistically, no performance criteria or pass rates are given, and the only timing experiment is a single profiling run (Fig. 8). There is no comparison to an optimal or reference planner and no optimality-gap measurement, and the fuzz test in Sec. V-A exercises partially defined target states rather than obstacle avoidance. The single moving-obstacle example uses constant motion. Additional randomized experiments with reported distributions, and a comparison against a global optimum or dense sampling, are needed before the headline claims can be accepted.","section":"Sec. V-B"}],"minor_comments":[{"comment":"The variable t in Eq. (10) is not defined in the text; please specify that it is the polynomial variable and describe how the root in the feasible interval is selected.","section":"Eq. (10)"},{"comment":"Section IV-A says obstacle motion can be described by concatenated polynomials of rank at most three, but Section V-B later says the method 'assumes a constant motion of the obstacle'; please clarify whether accelerating or otherwise non-constant obstacle motion is supported and evaluated.","section":"Sec. IV-A and Sec. V-B"},{"comment":"The text contains several typographical and formatting issues, including 'an time-optimal' in the Introduction, 'dont't' in the Related Work, and inconsistent spacing in 'MA V'; a careful proofread is needed.","section":"Throughout"},{"comment":"The statement that the code is open source points to a video/challenge webpage rather than a direct repository link; please provide the repository URL so the reproducibility claim can be verified.","section":"Sec. I"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' own previous trajectory generator [5], which is a normal line of research but should be reflected in the novelty framing: the main new algorithmic contribution is the viastate-insertion scheme, not the underlying trajectory optimization. The central issue for the editor is the mismatch between the unqualified time-optimality claims and the restricted, partly heuristic search family; this is fixable by substantially revising the claims and strengthening the evaluation, so I do not recommend rejection. Also, the open-source code link should be a direct repository for reproducibility."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The new thing here is real: extending the authors' prior time-optimal trajectory generator to partially defined target states, then using those to insert viastates that steer a triple-integrator MAV model around moving axis-aligned cuboids. The partial-target-state treatment is clean and the reported millisecond runtime on a single CPU core is plausible and well-profiled for the featured scenario. The open-source code is a plus. The fuzz test of 10,000,000 trajectories does show robustness for the trajectory generator, and the moving-obstacle collision check via polynomial zero-crossings is a sensible exploitation of the piecewise formulation.\n\nThe soft spot is exactly what the stress-test note flags. The abstract and conclusion say the method 'maintains time-optimality,' but Section IV-B.1 explicitly restricts viastates to obstacle corners with zero acceleration, and admits a nonzero acceleration could be slightly faster. That is not a fatal flaw — it is an honest engineering tradeoff — but it means the actual claim is 'optimal within a deliberately restricted family,' and the paper should say so in the abstract. The multi-velocity tradeoff in Section IV-B.3 is also approximate: quadratic polynomials fit through three points, solved by a quartic, so even within the corner/zero-acceleration family you are not guaranteed the exact optimum. The evaluation of the avoidance pipeline is only 100 scenarios with no statistics, no comparison to a global optimum, and no hardware experiments with dynamic obstacles. The fuzz test does not cover obstacle avoidance itself, only the partially defined target-state generator.\n\nThe paper is not circular; it builds on the authors' own [5] in an appropriate way, and the related work is fair. I would not call the central approach unsound, but the 'time-optimal' phrasing overreaches.\n\nWho is this for? Researchers working on fast online trajectory replanning for MAVs, especially those who want sub-10ms reactions to suddenly appearing obstacles. They will get a practical algorithm and a clear baseline, even if the optimality guarantee is conditional. A serious referee should engage with it — the contribution is real and the runtime claim is worth verifying — but the authors should be pushed to either prove the corner/zero-acceleration restriction is without loss of generality for the moving-obstacle case, or soften the claim and report the approximation gap empirically.\n\nRecommendation: send to peer review, with condition that the authors align the abstract with the actual theorem and add basic statistics over the scenarios.","headline":"A genuinely fast viastate-based avoidance pipeline with a real but qualified optimality claim; the paper itself concedes the 'time-optimal' label only holds under zero-acceleration corner viastates.","tokens_in":11713,"tokens_out":1248,"would_cite":false,"duration_ms":15958,"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":"This paper claims a planner that lets a fast-flying micro aerial vehicle dodge a suddenly perceived moving obstacle by inserting a viastate at an obstacle corner, while keeping the trajectory time-optimal and dynamically feasible.","keywords":["time-optimal trajectory","obstacle avoidance","micro aerial vehicle","viastate insertion","moving obstacles","bang-singular-bang trajectory","partial target state","real-time planning"],"falsifier":"Run a brute-force grid search over viastate velocities (as the paper itself does to visualize the cost surface) in a scenario where the obstacle is large and close to start or target, and compare the best grid trajectory with the planner's chosen trajectory; if any grid velocity yields a lower total time than the proposed root of the quadratic polynomials, the claim of time-optimality is disproved for that case.","tokens_in":10806,"feed_emoji":"🚁","tokens_out":5679,"duration_ms":56866,"temperature":0.7,"pith_summary":"This paper claims that a fast-flying micro aerial vehicle can dodge a suddenly perceived moving obstacle almost for free: the planner first builds a time-optimal, dynamically feasible trajectory to a partially specified target, checks it for collisions with axis-aligned cuboids, and when a collision is found inserts a viastate — an intermediate position, velocity, and acceleration — that routes the trajectory around a corner of the obstacle. The central assertion is that this insertion can be done while preserving time-optimality, and that the whole computation takes only a few milliseconds on one CPU core, fast enough for online replanning during aggressive flight. If true, the method closes a gap among existing planners: search-based approaches are slow or discretization-suboptimal, and optimization-based approaches take hundreds of milliseconds to seconds. The paper further claims the approach scales to any number of dimensions and handles obstacles that move with constant motion.","feed_headline":"Drones can dodge moving obstacles in under 6 ms","feed_subtitle":"It inserts corner-cutting viastates so the drone keeps a time-optimal path while evading cuboid obstacles.","key_machinery":"The central object is the viastate: a full state (position, velocity, acceleration) inserted between start and target, turning one trajectory into two segments that meet at an obstacle corner. The paper's key identity is the cost-surface structure of the two segments: for each bound axis, the time of a segment as a function of the viastate velocity has a single minimum and a plateau, so the total-time optimum is either the minimum of the dominant axis or the intersection of two quadratic polynomials fit through sampled points near the optimum. This, together with a database of closed-form solutions for the trajectory-shape equations (the 21 second-order conditions for the seven constant-jerk phases), is what lets the planner evaluate candidate trajectories in microseconds rather than milliseconds per sample.","core_discovery":"On the paper's own terms, the discovery is an algebraic shortcut for the otherwise expensive problem of choosing an optimal intermediate state. The authors model the MAV as a triple integrator with per-axis limits on velocity, acceleration, and jerk, and generate bang-singular-bang trajectories as concatenations of up to seven constant-jerk phases, with all closed-form solutions stored in a database. When the nominal trajectory collides with an axis-aligned cuboid, two 'bound' axes are chosen and the viastate position is fixed at the obstacle corner pushed out by the vehicle radius plus margin; the viastate acceleration is set to zero. The remaining degree of freedom is the two-element velocity vector at the viastate. The paper argues that the total time splits into two segments whose cost surfaces have characteristic 'grooves', and that the globally optimal velocity lies either at the extremum of a single dominant component or at the intersection of two quadratic polynomials fitted through three sample points per segment. This reduces what would require tens of thousands of trajectory evaluations to a handful of closed-form solves, with the whole avoidance pipeline, including collision checking, finishing in 5.98 ms in the reported scenario.","pith_inferences":["The paper leaves implicit that the zero-acceleration viastate is a safety-motivated restriction: it makes the vehicle's state at the closest approach easy to correct, but a small amount of acceleration toward the obstacle could shave off time; an extension would relax this constraint and compare the resulting speedup against the loss of robustness.","The quadratic-polynomial tradeoff is an approximation fitted through three sampled points; a natural test is to increase the number of samples or use the exact cost-surface curvature to see whether the intersection root is ever more than a few percent suboptimal.","The axis-aligned cuboid representation can be chained to represent more complex environments as unions of boxes, but the method as presented does not reason about overlapping or non-convex obstacles; extending the selection of bound axes to such unions is a direct next step.","Because the planner is so fast, it could be used as an inner loop inside a sampling-based or global planner, giving the global planner a fast dynamic-feasibility oracle rather than relying on a separate low-level controller."],"forward_implications":["A drone that perceives an obstacle suddenly can begin an avoidance maneuver within milliseconds, so the method can run as a low-level layer at typical control-loop frequencies.","Obstacles can be static or moving, and because the method treats each axis independently, the same pipeline works in four or more dimensions, e.g., adding yaw to handle constraints like 'do not look in direction X while inside volume Y'.","The number of candidate evasions grows as $p = n!/(n-2)!$ in $n$ dimensions, and each candidate is independent, so the computation parallelizes across cores.","Even though trajectories are time-optimal, the method can be used in a replanning loop: if the fastest candidate collides with another obstacle, the next fastest is selected, and uncertainty can be absorbed by increasing the obstacle margin and replanning frequently."],"supporting_citations":[{"why":"It supplies the full-state trajectory-generation method that this paper extends to obstacle avoidance.","marker":"[5]"},{"why":"It is the search-based minimum-time planner whose 3D runtime and discretization-dependent optimality serve as a baseline.","marker":"[11]"},{"why":"It extends the search-based planner to SE(3) and is used as another baseline for high runtime.","marker":"[12]"},{"why":"It is the sampling-based triple-integrator method with low runtime but no optimality guarantee.","marker":"[13]"},{"why":"It is the convex-optimization real-time planner with up to 81.4 ms runtime and a double-integrator model.","marker":"[14]"},{"why":"It is the chance-constrained MPC baseline that handles moving obstacles but does not report optimality.","marker":"[7]"},{"why":"It is the optimization-based planner that guarantees control-effort optimality but is slow.","marker":"[8]"},{"why":"It continues the optimization-based line, adding time allocation, again with runtime around one second.","marker":"[9]"},{"why":"It is the free-DoF trajectory method that produces minimum-snap paths via constrained gradient descent.","marker":"[15]"},{"why":"It provides third-order polynomial trajectories similar to the paper's but without obstacle avoidance.","marker":"[16]"}],"fun_headline_variants":["Algebraic shortcut lets drones dodge moving obstacles in 6 ms","Time-optimal dodge: viastate trick cuts replanning to milliseconds","MAV evasion: closed-form viastates enable 6 ms obstacle avoidance","Fast avoidance: algebraic groove solver for moving obstacles","Drone dodges cuboids in 5.98 ms with viastate shortcut"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claimed time-optimality rests on the assumption that the fastest evasive path belongs to the family of trajectories that pass through a corner of the obstacle with zero acceleration; if the true optimum accelerates through the corner, the generated trajectories are not time-optimal.","fun_headline_variants_meta":{"raw":{"variants":["Algebraic shortcut lets drones dodge moving obstacles in 6 ms","Time-optimal dodge: viastate trick cuts replanning to milliseconds","MAV evasion: closed-form viastates enable 6 ms obstacle avoidance","Fast avoidance: algebraic groove solver for moving obstacles","Drone dodges cuboids in 5.98 ms with viastate shortcut"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000398,"raw_usage":{"total_tokens":2081,"prompt_tokens":943,"completion_tokens":1138,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":559,"completion_tokens_details":{"reasoning_tokens":1042}},"tokens_in":559,"tokens_out":1138,"duration_ms":9288,"temperature":1.0,"reasoning_tokens":1042,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:55:46.176166+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a brute-force grid search over viastate velocities (as the paper itself does to visualize the cost surface) in a scenario where the obstacle is large and close to start or target, and compare the best grid trajectory with the planner's chosen trajectory; if any grid velocity yields a lower total time than the proposed root of the quadratic polynomials, the claim of time-optimality is disproved for that case.","supporting_citations":[{"cited_title":"Fast full state trajectory generation for multirotors,","cited_arxiv_id":null,"evidence_quote":"It supplies the full-state trajectory-generation method that this paper extends to obstacle avoidance."},{"cited_title":"Search-based motion planning for quadrotors using linear quadratic minimum time control,","cited_arxiv_id":null,"evidence_quote":"It is the search-based minimum-time planner whose 3D runtime and discretization-dependent optimality serve as a baseline."},{"cited_title":"Search-based motion planning for aggressive flight in SE(3),","cited_arxiv_id":null,"evidence_quote":"It extends the search-based planner to SE(3) and is used as another baseline for high runtime."},{"cited_title":"Aggressive 3-D collision avoidance for high-speed navigation,","cited_arxiv_id":null,"evidence_quote":"It is the sampling-based triple-integrator method with low runtime but no optimality guarantee."},{"cited_title":"Real-time quad- rotor path planning for mobile obstacle avoidance using convex optimization,","cited_arxiv_id":null,"evidence_quote":"It is the convex-optimization real-time planner with up to 81.4 ms runtime and a double-integrator model."},{"cited_title":"Chance-constrained collision avoidance for MA Vs in dynamic environments,","cited_arxiv_id":null,"evidence_quote":"It is the chance-constrained MPC baseline that handles moving obstacles but does not report optimality."},{"cited_title":"Quadrotor trajectory generation in dynamic environments using semi-definite relaxation on nonconvex QCQP,","cited_arxiv_id":null,"evidence_quote":"It is the optimization-based planner that guarantees control-effort optimality but is slow."},{"cited_title":"Optimal time allocation for quadrotor trajectory generation,","cited_arxiv_id":null,"evidence_quote":"It continues the optimization-based line, adding time allocation, again with runtime around one second."},{"cited_title":"Minimum snap trajectory generation and control for quadrotors,","cited_arxiv_id":null,"evidence_quote":"It is the free-DoF trajectory method that produces minimum-snap paths via constrained gradient descent."},{"cited_title":"Opening the door to new sensor-based robot applications - the reflexxes motion libraries,","cited_arxiv_id":null,"evidence_quote":"It provides third-order polynomial trajectories similar to the paper's but without obstacle avoidance."}],"review_version":1}