{"id":"b23d2b8e-5abb-4610-9716-0d38de8b7af7","arxiv_id":"2411.18086","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Time-varying buffered Voronoi cells and inter-visibility cells let a drone team plan collision-free, occlusion-free tracking trajectories in a distributed manner.","lead":"DMVC-Tracker is a distributed planning method that keeps a team of drones tracking a moving target while avoiding collisions and occluding one another's view. It introduces time-varying safety and visibility cells that move with the predicted target path, and the authors report millisecond-level computation in simulations and Crazyflie hardware tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DIVC occlusion avoidance is asserted without proof: inequality (11) is the load-bearing bridge from the cell half-spaces to no-occlusion, and the paper gives no derivation.","rationale":"I read the paper's central claim as a guarantee: DBVC and DIVC prevent mutual collisions and occlusions. The collision half has a real proof (Lemma 1 via Cauchy-Schwarz), and the reduced-conservativeness check in (21) is algebraically transparent. The simulations and hardware tests are supporting evidence, but they cannot certify a geometric guarantee. The load-bearing weak point is exactly the DIVC lemma: inequality (11) is the only link from the cell constraints checked in (20c) to the no-occlusion condition (4), and it is asserted rather than proved. The reader's weakest_assumption identifies this same point, so I agree. I add one concrete indication that the missing proof matters: at alpha=1, which the stated bounds allow, the strict inequality (11) fails in a tangent configuration, and the paper does not specify how alpha is chosen, so this equality case is not excluded by the planner. This does not overturn the conditional verdict; it is an addressable gap, not evidence that the method is fundamentally wrong. The 3D extension and the Table II formatting inconsistency are secondary and do not change the assessment.","tokens_in":13584,"tokens_out":22787,"duration_ms":218192,"concrete_test":"Run a global numerical falsification search for the DIVC lemma. Fix r_c=1 and sample 10^5 initial configurations (x_i0, x_j0) with target at the origin, distances in [1,10] m, and mutual angles covering [0, pi]. For each configuration, compute the feasible alpha interval from (10)/(14), select alpha at the lower bound, midpoint, and upper bound, construct the half-spaces (12)/(15), and solve the two global optimization problems min_{x_i in C_i, x_j in C_j, epsilon in [0,1]} ||epsilon x_i + (1-epsilon) x_q0 - x_j|| and the symmetric one, using branch-and-bound or dense grid sampling on the cell boundaries. If any minimum is <= r_c, the lemma is false and the central no-occlusion guarantee collapses.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DIVC prevents inter-agent occlusion rests entirely on Section IV-B. The planner only checks x_i(t) in the half-space intersection Hij_mu1(t) ∩ Hij_mu2(t) (Equation 20c); the claim that this implies the line-of-sight inequalities (11) for all pairs is stated as 'the following inequalities are satisfied' with no proof. Lemma 2 only shows the cells are nonempty by translating both agents with the target; it does not establish (11). The alpha bounds (10)/(14) are likewise justified only by asserted geometric conditions (points on segments, rays not intersecting). Because any trajectory passing (20c) lies in the cells, a single allowed configuration violating (11) would let two agents satisfy the DIVC check while one line of sight intersects the other agent. The strict form is already false at equality: with r_c=1, Q=(0,0), x_i=(2,0), x_j=(sqrt(3),1), the acute-case bound (14) admits alpha=1, the construction places both agents on the cell boundaries, and L(Q,x_i) is tangent to B(x_j,1), giving distance exactly r_c rather than > r_c. This concrete boundary case shows the missing derivation is not a formality; the paper also does not specify how alpha is chosen within the feasible range, so equality cases are not explicitly excluded.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed trajectory planner for multi-agent aerial target tracking. The key components are the Dynamic Buffered Voronoi Cell (DBVC) for inter-agent collision avoidance and the Dynamic Inter-Visibility Cell (DIVC) for inter-agent occlusion avoidance. These time-varying cells are combined with Bernstein-polynomial motion primitives and a sample-check-select strategy. The authors claim collision- and occlusion-free tracking with millisecond computation times, validated in simulations, hardware experiments, and two 3D scenarios.","tokens_in":13901,"tokens_out":6808,"duration_ms":59703,"significance":"If the central geometric claims are correct, the work has clear practical value: it offers a fully distributed, non-priority-based method for multi-agent tracking that simultaneously addresses collision, occlusion, and target-distance constraints, with a less conservative feasibility check than the authors' prior work. The DBVC Lemma 1 is proved, the use of Bernstein convex-hull properties is sound, and the empirical validation is extensive, including hardware demonstrations and comparisons against noncooperative and conservative baselines. The main unresolved issue is that the DIVC occlusion-avoidance property, which is the load-bearing theoretical bridge between the half-space constraints and the no-occlusion guarantee, is asserted without proof.","major_comments":[{"comment":"If the authors cannot provide such a proof, the claim that DIVC prevents inter-agent occlusion is unsupported, and the validation results would not justify the proposed method's main contribution.","section":"Section IV-B, Eqs. (10)-(16) and Eq. (20c)"},{"comment":"As written, the 3D experimental results cannot be interpreted as evidence for the same collision- and occlusion-free guarantee that is claimed for the 2D case.","section":"Section VI-D, 3D Extensions"},{"comment":"The authors should also verify that the same ordering convention is used across all rows and columns of the table.","section":"Table II, Scenario 1 row"}],"minor_comments":[{"comment":"This notational inconsistency makes the construction harder to follow.","section":"Section IV-B, acute case"},{"comment":"Please correct the subscript for consistency with the other sigma terms.","section":"Section V-B-2, Eq. (21)"},{"comment":"Clarify that both agents translate with the target (or otherwise specify what is assumed about the other agent's motion) so that the non-emptiness argument is complete.","section":"Section IV-A, Lemma 2"},{"comment":"This is a minor language issue that does not affect the technical content.","section":"Section VI-A, Scenario 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of IEEE Robotics and Automation Letters, and the practical results are potentially of interest. The main risk is the unproved DIVC occlusion-avoidance property; this is not a question of style but of the central guarantee. If the authors can supply a rigorous proof, or a modified construction with a proof, and fix the Table II inconsistency, the paper could become acceptable. The reuse of the authors' previous work [8] is substantial but appears to be appropriately acknowledged; the novelty lies in the DBVC/DIVC cells and the less conservative feasibility checks."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: useful paper, real new idea, but the load-bearing occlusion-avoidance claim is not actually proven. The DBVC and DIVC are time-varying cells that move with the target, a genuine step beyond the static BVC and the authors' single-agent BPMP tracker. Lemma 1 for collision avoidance is fine. The less conservative Bernstein-based feasibility check is a real improvement, and the validation is extensive: simulations with dozens of obstacles, Crazyflie hardware, and success-rate comparisons.\n\nThe soft spot is Section IV-B. Inequality (11) is the entire bridge from the DIVC half-space constraints to 'no agent occludes another agent's line of sight.' The paper states it as 'the following inequalities are satisfied' and moves on. Lemma 2 only shows the cells are nonempty by translating with the target; it does not establish (11). The alpha bounds in (10)/(14) are justified by geometric claims, but no derivation is given. That matters: the strict form fails at equality in at least one concrete configuration (the stress-test example), and the paper never tells you how alpha is chosen inside the feasible range, so equality cases are not ruled out. This gap is not fatal if patched; the rest of the paper is assembled carefully, and I would expect a proof to fit in an appendix. Right now, the main guarantee rests on an unproved lemma.\n\nSmaller issues: the 3D extension is only examples, no mathematical generalization—fine for a letter as long as it is described as heuristic. Table II has at least one row that looks internally inconsistent (Sc1 proposed row gives agent-to-agent distance min/mean 0.561/0.727 and line-of-sight-to-agent distance min/mean 0.290/0.727; the identical means with different minima are at least suspicious). No code or parameters are released, which makes the success-rate tables hard to audit.\n\nFor whom: anyone in multi-drone tracking or distributed trajectory generation should read this; it is within-subfield significant. It deserves a serious referee. I would send it to review with the expectation that the DIVC proof appears or the claim is softened. The central idea is plausible and the flaw is addressable.","headline":"A genuinely new distributed-tracking construction with a real gap: the DIVC occlusion guarantee is asserted, not proved, and the paper needs a fix before its central claim is solid.","tokens_in":14412,"tokens_out":2977,"would_cite":true,"duration_ms":25275,"reading_group":"yes","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 that two time-varying cells—a Dynamic Buffered Voronoi Cell and a Dynamic Inter-Visibility Cell—make distributed multi-agent target tracking collision-free, occlusion-free, and computable in milliseconds.","keywords":["distributed multi-agent planning","target tracking","occlusion avoidance","collision avoidance","dynamic Voronoi cells","Bernstein polynomial motion primitives","aerial robotics","trajectory planning"],"falsifier":"Numerically test inequality (11) for a dense grid of agent and target positions near the boundary of the allowed $\\alpha$ ranges (10) and (14); if any point in the DIVC intersection lies within $r_c$ of the opposite line-of-sight segment, the core occlusion guarantee is false. A simpler experiment is to run the planner with two agents in an empty environment and measure the minimum distance between each line of sight and the other agent over time; a value below $r_c$ falsifies the claim.","tokens_in":13362,"feed_emoji":"🛸","tokens_out":5328,"duration_ms":43579,"temperature":0.7,"pith_summary":"This letter claims that two time-varying geometric cells can make distributed multi-agent target tracking both collision-free and occlusion-free in real time. The Dynamic Buffered Voronoi Cell (DBVC) keeps agents apart, while the Dynamic Inter-Visibility Cell (DIVC) keeps each agent's line of sight to the target clear of its teammates. Because the cells translate with the target's predicted motion, agents can plan independently in milliseconds, which the authors demonstrate in simulations and hardware flights with dozens of obstacles.","feed_headline":"Moving safety cells keep drone teams collision- and occlusion-free","feed_subtitle":"Dynamic Voronoi and visibility cells let each drone plan in milliseconds while keeping the target in view.","key_machinery":"The load-bearing objects are the Dynamic Buffered Voronoi Cell (DBVC), a time-shifting half-space that guarantees inter-agent separation, and the Dynamic Inter-Visibility Cell (DIVC), a pair of half-spaces that are supposed to keep each agent's view of the target unblocked by teammates. The DBVC generalizes the static Buffered Voronoi Cell by translating with the target's predicted displacement $x_q(t)-x_{q0}$, which prevents agents from getting stuck while following a moving target. The DIVC is built in two cases depending on whether the angle between the two agents' lines of sight is obtuse or acute; in each case the half-spaces are defined so that, for a parameter $\\alpha$ within the allowed range, any point in their intersection satisfies the line-of-sight clearance inequality (11). Bernstein polynomial motion primitives carry the trajectories, and their convex hull property lets the planner verify inclusion in these affine cells by checking control points, while their integral property speeds up the selection of the best primitive.","core_discovery":"The central claim is that the DBVC and the DIVC are time-variant spaces that prevent mutual collisions and occlusions among agents while enabling them to maintain suitable distances from the moving target. The DBVC is a half-space built from the agents' current positions and the target's predicted trajectory, and Lemma 1 shows that if each of two agents lies in its own DBVC half-space, their distance is at least $2r_c$. The DIVC is the intersection of two half-spaces per agent pair, constructed separately for obtuse and acute viewing angles, and the paper asserts that any trajectory inside them keeps the segment from the agent to the target at distance greater than $r_c$ from every other agent. These cells are combined with Bernstein polynomial motion primitives and a sample-check-select strategy; the convex hull and integral properties of Bernstein polynomials make the feasibility checks fast, and a refined, less conservative check than the authors' previous work increases the success rate. The algorithm computes each agent's trajectory within several milliseconds on an Intel i7 desktop.","pith_inferences":["A proof or counterexample for inequality (11) would settle whether the occlusion-free guarantee is absolute; the simulations alone cannot distinguish a valid cell construction from a plannable one that occasionally fails.","The cell construction assumes synchronized replanning and shared positions; asynchronous communication or latency would break the geometric assumptions, so the method's claims are strongest for tightly synchronized teams.","The same translated-cell idea could be extended to shape-changing cells (the authors list this as future work) or to heterogeneous agent radii, which would let the planner handle mixed fleets.","A direct extension of the DIVC to 3D would require the obtuse/acute construction to be replaced by a spherical or conic condition, so the 3D demonstrations shown here are empirical rather than guaranteed by the lemmas."],"forward_implications":["Each agent can replan independently from shared positions and the predicted target trajectory, without priority ordering or consensus among teammates.","The time-varying cells avoid the deadlock that static buffered Voronoi cells cause when the target moves, keeping agents within the commanded tracking distance.","The relaxed feasibility check expands the set of admissible primitives, which the benchmark shows as higher success rates in dynamic obstacle fields.","Reported per-agent computation under 10 ms at 1000 sampled primitives supports real-time replanning on small onboard computers.","In the comparison, the DBVC and DIVC raise success rates dramatically as the number of trackers grows, especially at short tracking distances."],"supporting_citations":[{"why":"Supplies the static Buffered Voronoi Cell that the DBVC extends into a time-varying form.","marker":"[7]"},{"why":"Provides the Bernstein polynomial motion primitive trajectory generation, the base feasibility checks that the paper refines, and the target trajectory prediction method.","marker":"[8]"},{"why":"The state-of-the-art multi-drone planner used as the baseline in the unstructured-environment comparison.","marker":"[20]"},{"why":"Generates the pre-calculated paths followed by the target and dynamic obstacles in the hardware experiments.","marker":"[21]"}],"fun_headline_variants":["Dynamic Voronoi and visibility cells enable millisecond drone tracking","Distributed drones track targets with buffered Voronoi safety cells","Millisecond multi-agent trajectory planning for target tracking","Dynamic safety cells keep drone teams collision-free and tracking","Fast distributed multi-agent tracking using dynamic Voronoi cells"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's occlusion-avoidance guarantee depends on an unproved geometric assertion: that for any $\\alpha$ in the stated range, every point in the DIVC intersection of two half-spaces lies at distance greater than $r_c$ from the line segment joining the other agent to the target; if that assertion fails for some configuration, the method can produce trajectories that block a teammate's view even though the planner ran exactly as specified.","fun_headline_variants_meta":{"raw":{"variants":["Dynamic Voronoi and visibility cells enable millisecond drone tracking","Distributed drones track targets with buffered Voronoi safety cells","Millisecond multi-agent trajectory planning for target tracking","Dynamic safety cells keep drone teams collision-free and tracking","Fast distributed multi-agent tracking using dynamic Voronoi cells"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000811,"raw_usage":{"total_tokens":3532,"prompt_tokens":893,"completion_tokens":2639,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":509,"completion_tokens_details":{"reasoning_tokens":2558}},"tokens_in":509,"tokens_out":2639,"duration_ms":16778,"temperature":1.0,"reasoning_tokens":2558,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:32:23.478377+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerically test inequality (11) for a dense grid of agent and target positions near the boundary of the allowed $\\alpha$ ranges (10) and (14); if any point in the DIVC intersection lies within $r_c$ of the opposite line-of-sight segment, the core occlusion guarantee is false. A simpler experiment is to run the planner with two agents in an empty environment and measure the minimum distance between each line of sight and the other agent over time; a value below $r_c$ falsifies the claim.","supporting_citations":[{"cited_title":"Fast, on-line collision avoidance for dynamic vehicles using buffered voronoi cells,","cited_arxiv_id":null,"evidence_quote":"Supplies the static Buffered Voronoi Cell that the DBVC extends into a time-varying form."},{"cited_title":"Bpmp-tracker: A versatile aerial target tracker using bernstein polynomial motion primitives,","cited_arxiv_id":null,"evidence_quote":"Provides the Bernstein polynomial motion primitive trajectory generation, the base feasibility checks that the paper refines, and the target trajectory prediction method."},{"cited_title":"Swarm of micro flying robots in the wild,","cited_arxiv_id":null,"evidence_quote":"The state-of-the-art multi-drone planner used as the baseline in the unstructured-environment comparison."}],"review_version":1}