{"id":"4ffb303e-ea56-4be6-a46f-d405cc790642","arxiv_id":"2412.20695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"CoCap adapts conflict-based search to coordinate multi-drone camera views for outdoor motion capture, improving coverage rewards over sequential planning in simulated cluttered scenes.","lead":"When several camera drones film moving actors outdoors, planning each drone's views in sequence can produce poor coverage and conflicts. This paper adapts a coordination method from multi-robot pathfinding so the drones negotiate their camera views, and reports that the approach beats sequential planning in two obstacle-heavy simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim rests on an unvalidated pixel-density proxy and an unquantified unconstrained baseline, so the reported reward gains do not yet establish 'approaches ideal outcomes'.","rationale":"Reading the paper in good faith, the central contribution is algorithmic: a CBS-style constraint tree for multi-camera view planning, evaluated by the GreedyPerspectives reward. For the abstract's claim to hold, the reward metric must be a faithful proxy for motion-capture quality, and the baseline comparison must be complete. The metric is the least secure link. The authors themselves list the simplifications that break the link (cuboid actors, 2.5D height map, no real deployment), and no reconstruction-quality evaluation is provided. This is not an external-disagreement objection; it is an internal-evidence objection: the paper's own stated limitations identify the missing validation. The missing no-constraint baseline total in Table I, together with the weak bottleneck margin and the below-baseline search variant, makes the headline claim stronger than the reported numbers. I do not see a clear internal inconsistency in the CBS adaptation; the low-level search is explicitly approximate, and the best-first return policy is a reasonable design choice. The paper deserves credit for two concrete scenarios and for reporting compute time in the corridor case, but the conditional verdict should remain until the reward proxy is tested against an actual pose-reconstruction metric.","tokens_in":8137,"tokens_out":9287,"duration_ms":101133,"concrete_test":"Render the camera trajectories produced by CoCap (MDP) and Sequential (MDP) in the corridor scenario with a photorealistic renderer and feed the resulting views into a multi-view 3D pose estimator; compare pose error or reconstruction completeness. If CoCap's reward advantage does not translate into lower pose error, the abstract's claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CoCap 'approaches the ideal outcomes of unconstrained planning' and 'outperforms existing sequential planning methods' in occluded, obstacle-rich scenes. The only evidence offered is the pixel-density reward in Eq. (1), measured in two deterministic simulated scenarios. This is load-bearing for two reasons. First, Table I never reports the total reward of the no-inter-robot-constraint baseline that is the 'ideal outcome', so the gap CoCap is supposed to close is never quantified. Second, the reported advantage is fragile: CoCap (MDP) beats Sequential (MDP) by only 5162 vs 5053 in the bottleneck scenario, and the CoCap (Search) variant scores 3922 in the corridor, below Sequential (MDP) at 4127, yet the abstract states 'outperforming sequential planning methods' without qualification. The paper's own limitations paragraph concedes that actors are cuboids, occlusion is modeled by a 2.5D height map, and real multi-UAV deployment has not been tested. If the pixel-density reward rewards viewpoint configurations that do not actually improve 4D pose reconstruction in cluttered scenes, the central claim does not follow from the reported numbers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CoCap, a coordinated multi-UAV view-planning approach for motion capture of multiple actors in cluttered outdoor environments. CoCap applies Conflict-Based Search (CBS) principles: a greedy joint plan is first computed without inter-robot constraints; when robot-robot conflicts are detected, constraints are added and the affected robots are replanned with a low-level single-agent view planner. The paper also presents a fast heuristic search variant for single-agent planning. Experiments in two simulated scenarios (corridor, bottleneck) report that CoCap with value iteration achieves higher total pixel-density reward than sequential planning (4662 vs 4127 in the corridor; 5162 vs 5053 in the bottleneck), while the search variant is much faster but scores lower (3922 in the corridor). The paper concludes that CoCap approaches the performance of unconstrained planning and outperforms sequential planning.","tokens_in":8328,"tokens_out":5979,"duration_ms":57446,"significance":"If the reported results are taken at face value, CoCap is a plausible approach for coordinating multiple camera-equipped robots in obstacle-rich environments, and the combination of CBS-style conflict resolution with a coverage objective is a reasonable extension of prior work. The single-agent search variant addresses real-time computation. The paper is honest in its limitations paragraph about cuboid actor models, 2.5D height maps, and absence of real-world deployment. However, the evidence is far from establishing the central claims: only two deterministic scenarios, no error bars, no reported unconstrained baseline totals, a reward surrogate that is not validated against actual pose reconstruction, and one variant underperforming the sequential baseline. These gaps mean the current version is not yet ready for archival publication.","major_comments":[{"comment":"The unconstrained (\"no inter-robot constraint\") baseline reward is never reported numerically; Fig. 5 plots its cumulative reward but Table I omits it. Since the abstract claims that CoCap \"approaches the ideal outcomes of unconstrained planning,\" the gap between CoCap's totals (4662, 5162) and the unconstrained totals must be quantified. Please report the unconstrained total reward for both scenarios and state the fraction of the gap recovered by CoCap.","section":"Table I / Sec. VI-A"},{"comment":"All performance numbers in Table I and Fig. 5 come from a single deterministic run per scenario, with no repeated trials or variance measures. The corridor advantage is 4662 vs 4127, and the bottleneck advantage is only 5162 vs 5053; without error bars or multiple random initial positions, it is impossible to know whether these differences are robust. Please provide statistics over multiple runs (e.g., random robot starts or actor trajectories).","section":"Table I / Fig. 5"},{"comment":"CoCap (Search) achieves 3922 in the corridor, which is lower than Sequential (MDP) at 4127; the abstract nevertheless states without qualification that CoCap \"outperforms existing sequential planning methods.\" Please either qualify the claim to \"CoCap with value iteration\" or explain why the search variant's lower reward is acceptable (e.g., it is a different trade-off point). Also, the bottleneck row lacks CoCap (Search) and all compute-time entries for the bottleneck scenario, making the comparison incomplete.","section":"Abstract / Table I, corridor row"},{"comment":"The evaluation uses the pixel-density coverage reward from GreedyPerspectives as the sole measure of performance. The paper's own limitations paragraph concedes that actors are modeled as cuboids, occlusion is via a 2.5D height map, and real multi-UAV deployment has not been tested. As a result, the abstract's claim about \"motion capture\" quality rests on an unvalidated surrogate. Please either provide evidence that the reward correlates with 4D pose reconstruction accuracy (e.g., running a pose estimator on rendered views) or restrict the claims to coverage reward.","section":"Sec. VI / Eq. (1)"},{"comment":"The algorithms are under-specified and contain apparent inconsistencies. In Algorithm 2, line 8, the update \"Rk+1 ← (Rk+1 + Rk)γk−1\" references R_{k+1} before assignment, and the discounting is ambiguous; the branching factor ST, the discount factor γ, and the exact definition of availableActions (including obstacle and motion constraints) are not given. In Algorithm 1, the tree ordering is said to expand the node with max g, but CBS expands by least cost; the paper must state clearly that TREE is a max-priority queue and define tie-breaking. These details are needed to reproduce the reported results.","section":"Algorithms 1 and 2 / Sec. V"}],"minor_comments":[{"comment":"There is a typo: \"GreedyPrespectives\" should be \"GreedyPerspectives\".","section":"Sec. I"},{"comment":"The notation \"g = J(Ξ)\" is introduced without defining the function J or the symbol Ξ; please define these symbols in the problem formulation.","section":"Sec. III"},{"comment":"The curves in Fig. 5 are not labeled in the figure itself, and the captions read \"Scale rewards\" (likely \"Scaled rewards\"). Please add legends and explain the scaling.","section":"Fig. 5"},{"comment":"The constraint ω_i = (i, u_i^a, u_i^b, t) should be explicitly defined as forbidding agent i from traversing the edge (u_i^a, u_i^b) at time t, rather than leaving it to the reader's inference.","section":"Sec. IV"},{"comment":"The expression \"Px \\ {Pxitk}\" uses undefined subscripts i, t, k; please define this set-minus notation or replace it with a clearer description.","section":"Sec. V-A"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is more in the style of a conference paper than a full journal article: the central idea is interesting, but the evidence is preliminary and the presentation is under-specified. The strongest concern is that the performance claims outrun the experimental support. I would suggest the editors consider whether the paper, in its current form, meets the journal's bar for archival publication; the proposed revisions would substantially strengthen it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is applying Conflict-Based Search to camera-view planning rather than just path planning. Previous multi-drone view planners use sequential or prioritized planning; CoCap instead builds a constraint tree over robot trajectories and re-plans individual agents to recover coverage reward lost to collision constraints. That is a real adaptation of CBS, and the paper clearly explains the algorithmic changes: a coverage reward as the node cost, greedy initialization, and a value-iteration low-level solver. The single-agent search variant is also useful, cutting compute from hundreds of seconds to 2.7 seconds, even if at some reward cost. Credit where due: the problem is well motivated, the formalization is careful, and the limitations paragraph is unusually candid about the cuboid actors, the 2.5D height map, the lack of real deployment, and the centralized-communication risk.\n\nThe soft spots are concentrated in the evidence for the central claim. Table I reports only two deterministic scenarios with no error bars, no repeated runs, and no code or data released. The abstract says CoCap \"outperforms existing sequential planning methods,\" but in the corridor the Search variant scores 3922 versus Sequential-MDP's 4127, and the bottleneck improvement is only 5162 versus 5053. The unconstrained baseline, the \"ideal outcome\" CoCap is said to approach, never appears as a number in Table I; it only shows up in figures, so the gap being closed is never quantified. And the reward itself, inherited from GreedyPerspectives, is a pixel-density proxy that nobody has validated against actual 4D pose reconstruction quality. These are not fatal flaws—they are addressable with randomized trials, released artifacts, and a table that includes the unconstrained baseline—but they do mean the paper's strongest sentence goes beyond what the reported numbers support.\n\nThe citation pattern is fine. Reusing the GreedyPerspectives reward is not circular because the comparison is between coordination mechanisms under the same reward, not a fit to a prior result. The related work is honest about CBS being underexplored in perception planning.\n\nWho is this for? Researchers working on multi-drone perception, especially anyone building on view-planning or MAPF. It does not resolve a long-standing scientific question, but it is a solid engineering contribution that deserves a careful referee rather than a desk reject. My recommendation: send it to review, but push the authors to report error bars, add the unconstrained baseline to the table, and qualify the outperform claim to match the actual scenario-level results.","headline":"A legitimate CBS-for-view-planning extension that works in two simulated scenarios, but the headline claim outruns the evidence.","tokens_in":8876,"tokens_out":1318,"would_cite":true,"duration_ms":15519,"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":"Coordinated conflict-based view planning lets multi-drone camera teams capture cluttered outdoor scenes nearly as well as if the drones never had to avoid one another.","keywords":["multi-robot view planning","motion capture","aerial cameras","conflict-based search","coverage optimization","occlusion-aware planning","UAV coordination","multi-actor scenes"],"falsifier":"Record the corridor and bottleneck scenarios with a photorealistic renderer or real drones, run CoCap and sequential planning on the same actor trajectories, and compare the resulting 4D pose reconstruction error (for example mean per-joint 3D error) rather than the coverage reward. If sequential planning produces equal or lower pose error despite lower reported coverage, the paper's central claim is falsified.","tokens_in":7913,"feed_emoji":"🎥","tokens_out":7385,"duration_ms":69883,"temperature":0.7,"pith_summary":"This paper tries to establish that a team of camera-equipped drones can capture multiple moving actors in cluttered outdoor scenes almost as well when they must avoid each other as when they are allowed to fly freely. The proposed approach, CoCap, adapts conflict-based search from multi-agent navigation to view planning: drones initially plan greedily, and only when two trajectories collide does the planner add a constraint and re-plan one drone, repeating until a conflict-free joint plan is found. In two simulated scenarios with obstacles, narrow passages, and heavy occlusion, CoCap reports higher total coverage reward than sequential priority-based planning and approaches the reward of unconstrained planning. The result matters because real outdoor motion capture needs multiple coordinated viewpoints for 4D pose reconstruction, and collision constraints should not force drones to give up those viewpoints.","feed_headline":"Conflict-first planning recovers drone camera coverage in clutter","feed_subtitle":"CoCap adds collision constraints only when drones conflict, preserving multi-actor view rewards that sequential planners lose.","key_machinery":"The central mechanism is a constraint tree whose nodes hold a joint robot trajectory, a set of inter-robot constraints, the accumulated pixel coverage over actor faces, and the corresponding reward. When the pairwise conflict check $\\Psi$ finds a collision between two drones, the node splits into two children, each adding a constraint to one of the two drones and re-planning that drone with the coverage already collected by the other drones held fixed. The objective is the incremental coverage gain of Eq. (1), the sum over actor faces of square-root marginal pixel-density gains, which rewards diverse multi-view coverage. A low-level single-agent view search, guided by a coverage heuristic and a discount factor, replaces the expensive value-iteration solver and allows fast replanning under the new constraints.","core_discovery":"CoCap claims that the main loss of coverage in multi-drone motion capture comes from fixing a priority order for drones and planning them one after another, so that later drones inherit all earlier constraints. By instead treating conflicts as they appear and re-planning only the involved drone, CoCap keeps the rewards of diverse viewpoints while ensuring inter-robot safety. Concretely, in the corridor scenario CoCap reaches a total scaled reward of 4662 against 4127 for sequential planning, and in the bottleneck scenario 5162 against 5053; in both cases it approaches the no-constraint upper bound. The paper also claims a single-agent view search that plans in about 2.7 seconds compared with 3981 seconds for value iteration, at a reward cost of 3922 versus 4662, making online replanning feasible.","pith_inferences":["The paper does not report how the constraint tree grows with more than four drones; a natural test is scaling to eight or more drones to see whether the first-conflict-free solution remains computationally feasible.","If the pixel-density proxy holds, the same constraint-tree coordination could be applied to other multi-camera perception tasks, such as outdoor 3D reconstruction or surveillance, by swapping in their coverage rewards.","The single-agent search's lack of lookahead suggests a testable extension: adding a limited-depth rollout for later high rewards could raise its reward without sacrificing its speed advantage.","Real deployment would need to handle communication latency and centralization; a decentralized variant that resolves local conflicts without a central tree is a direct next step that the paper leaves unexplored."],"forward_implications":["In narrow, obstacle-dense scenes, CoCap closes most of the gap between sequential planning and the unconstrained upper bound, so collision avoidance need not be the dominant cost of multi-drone coverage.","CoCap's reward advantage appears in both tested scenarios, and the gap between CoCap and sequential planning widens while the drones are inside the cluttered region where conflicts are frequent.","The single-agent view search reduces planning time from thousands of seconds to a few seconds, which makes online replanning practical for real-time capture, with a measurable but modest reward loss.","Coordinating view planning through a constraint tree, rather than a fixed priority order, is shown to be a viable way to enforce inter-robot constraints in perception tasks.","The reported comparisons use the same coverage reward for all methods, so the improvement is attributed to the coordination mechanism rather than to a different objective."],"supporting_citations":[{"why":"Supplies the pixel-density coverage reward (Eq. 1) and the sequential planning baseline CoCap is compared against.","marker":"[7]"},{"why":"Provides the conflict-based search constraint-tree mechanism that CoCap adapts from multi-agent navigation to view planning.","marker":"[11]"},{"why":"Introduces the multi-actor pixel-density coverage model over polygonal actor approximations that CoCap extends with obstacle and occlusion handling.","marker":"[13]"},{"why":"Establishes single-actor outdoor aerial motion capture with collaborative cameras, the scenario CoCap generalizes to coordinated multi-actor scenes.","marker":"[6]"},{"why":"Represents the reinforcement-learning aerial capture approach that handles single actors but lacks obstacle reasoning, marking the gap CoCap targets.","marker":"[12]"}],"fun_headline_variants":["CoCap: conflict-aware drone planning beats sequential capture","Drone mocap with conflict-first planning keeps multi-view coverage","CBS-style coordination boosts outdoor multi-drone capture quality","Conflict-triggered replanning preserves drone view diversity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim depends on the assumption that the pixel-density coverage reward computed over cuboid actor faces is a faithful proxy for real motion-capture quality, and that a 2.5D height map correctly models occlusion; if either fails, the simulated reward advantage may not translate into better real-world pose reconstruction.","fun_headline_variants_meta":{"raw":{"variants":["CoCap: conflict-aware drone planning beats sequential capture","Drone mocap with conflict-first planning keeps multi-view coverage","CBS-style coordination boosts outdoor multi-drone capture quality","Conflict-triggered replanning preserves drone view diversity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000587,"raw_usage":{"total_tokens":2710,"prompt_tokens":849,"completion_tokens":1861,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":465,"completion_tokens_details":{"reasoning_tokens":1793}},"tokens_in":465,"tokens_out":1861,"duration_ms":12489,"temperature":1.0,"reasoning_tokens":1793,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T23:13:08.737513+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Record the corridor and bottleneck scenarios with a photorealistic renderer or real drones, run CoCap and sequential planning on the same actor trajectories, and compare the resulting 4D pose reconstruction error (for example mean per-joint 3D error) rather than the coverage reward. If sequential planning produces equal or lower pose error despite lower reported coverage, the paper's central claim is falsified.","supporting_citations":[{"cited_title":"Greedy per- spectives: Multi-drone view planning for collaborative percep- tion in cluttered environments,","cited_arxiv_id":null,"evidence_quote":"Supplies the pixel-density coverage reward (Eq. 1) and the sequential planning baseline CoCap is compared against."},{"cited_title":"Conflict-based search for optimal multi-agent pathfinding,","cited_arxiv_id":null,"evidence_quote":"Provides the conflict-based search constraint-tree mechanism that CoCap adapts from multi-agent navigation to view planning."},{"cited_title":"Multi- robot planning for filming groups of moving actors leveraging submodularity and pixel density,","cited_arxiv_id":null,"evidence_quote":"Introduces the multi-actor pixel-density coverage model over polygonal actor approximations that CoCap extends with obstacle and occlusion handling."},{"cited_title":"3D Human Reconstruction in the Wild with Collaborative Aerial Cameras","cited_arxiv_id":"2108.03936","evidence_quote":"Establishes single-actor outdoor aerial motion capture with collaborative cameras, the scenario CoCap generalizes to coordinated multi-actor scenes."},{"cited_title":"Aircaprl: Autonomous aerial hu- man motion capture using deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Represents the reinforcement-learning aerial capture approach that handles single actors but lacks obstacle reasoning, marking the gap CoCap targets."}],"review_version":1}