{"id":"f46d6eb0-6a7f-41db-b99a-17a260d14b3e","arxiv_id":"2603.24238","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"PSTO, an egocentric LiDAR-plus-intent grid, lets a single decentralized MARL policy capture faster than privileged SOTA and fly zero-shot outdoor multi-drone pursuit.","lead":"A drone swarm learns to chase an evader through clutter by feeding raw LiDAR plus predicted teammate and target motion into one shared neural policy. The work shows faster captures than privileged-state baselines and a real outdoor 2-vs-1 flight with only onboard computers.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Shared teammate/evader kinematics (via comms and Swarm-LIO proxy) mean PSTO’s intent channel is not pure local raw-LiDAR, so the end-to-end “only onboard sensing” claim is only partially demonstrated.","rationale":"The reader’s weakest_assumption correctly isolates the shared-state dependency that undercuts the pure “raw-LiDAR / only-onboard” framing. Under the paper’s stated hardware-aligned assumptions the empirical results remain convincing: PSTO still beats privileged OPEN on capture time, the ablations isolate the value of spatial alignment, scalability holds across team sizes, and the outdoor 2-vs-1 flight is non-trivial. No deeper internal inconsistency or unaddressed failure mode appears; the limitation is already flagged for future fully-onboard detection. Therefore the CONDITIONAL verdict (accept once perception is fully local and artifacts are released) needs no adjustment.","tokens_in":14822,"tokens_out":566,"duration_ms":29171,"concrete_test":"Re-evaluate the released (or re-trained) policy on the exact 2-vs-1, 1.6 m/s, 9-obstacle condition of Table I after injecting realistic communication degradation: 20–50 % packet loss plus Gaussian noise (σ matching reported Swarm-LIO residuals) into shared teammate states and the effective evader position of Eq. 4. If success rate falls below ~85 % or capture time rises by more than 15 steps relative to the perfect-sharing baseline, the outdoor and end-to-end claims require explicit qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that a single decentralized policy maps the two-channel PSTO (raw LiDAR proximity + predictive intent heatmap) directly to velocity commands, yielding superior capture efficiency versus privileged baselines and outdoor transfer with only onboard sensing/compute. Yet Sec. I and III-B-iii explicitly feed perfect teammate world-frame positions and velocities over a communication network into linear extrapolation (Eq. 8) to populate H_Teammate; evader observability is likewise treated as shared (“if observable by any pursuer,” Eq. 4) and outdoor runs (Sec. IV-E) replace onboard detection with shared Swarm-LIO estimates. Consequently the H_intent channel is not generated from local raw sensors alone. If those links are lost, delayed or noisy, the spatially-aligned representation that underpins both the efficiency gains and the zero-shot outdoor result degrades, and the pure perception-to-control framing no longer holds. This is the least secure premise for the strongest claim.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a decentralized MARL framework for multi-AAV pursuit–evasion in cluttered arenas that maps a two-channel egocentric representation—Predictive Spatio-Temporal Observation (PSTO)—to continuous body-frame velocity commands. PSTO projects raw LiDAR ranges into a fixed-resolution spherical proximity map and overlays a predictive intent heatmap built from an LSTM forecast of the evader and linear extrapolations of teammate trajectories (Eqs. 3–11). Under CTDE/MAPPO with a shaped reward and progressive curriculum, a single shared policy is trained to navigate obstacles, intercept a faster APF-controlled evader, and maintain encirclement. Simulations (Table I, 150 rollouts) report competitive success rates and shorter capture times than traditional heuristics and the privileged-state OPEN baseline, plus zero-shot scaling across 2/3/4-vs-1 teams (Table II). Fully autonomous outdoor 2-vs-1 flights on quadrotors with onboard LiDAR and NUC compute are presented as zero-shot validation.","tokens_in":15109,"tokens_out":1740,"duration_ms":28259,"significance":"If the claims hold under the paper’s stated sensing and communication assumptions, the work is a solid systems-level contribution to multi-UAV pursuit: it is among the first to close a dense LiDAR-to-control loop for cooperative aerial P–E rather than operating on hand-crafted geometric features or ground-truth obstacle maps. Strengths that should be credited include (i) a carefully designed ablation suite isolating cooperation, intent, and spatial alignment (Table I; Fig. 4), (ii) a curriculum that yields a single policy that generalizes across team sizes without retraining (Table II), and (iii) outdoor hardware transfer with measured onboard latency (~3.87 ms). The spatial unification of dense geometry and sparse predictive intent in a fixed-resolution grid is a reusable representation idea for other multi-agent aerial tasks. The result is incremental rather than foundational, but it is practically relevant for infrastructure-free swarm deployment.","major_comments":[{"comment":"The abstract and Sec. I repeatedly frame the method as mapping “raw LiDAR observations directly” to control and as relying on “only onboard sensing and computing.” This is not accurate for the full observation. Sec. III-B-(iii) and Eq. (8) inject perfect world-frame teammate positions and velocities obtained over a communication network into H_Teammate; Eq. (4) treats evader observability as shared (“if observable by any pursuer”); and Sec. IV-E replaces onboard evader detection with shared Swarm-LIO estimates as an “experimental simplification.” Consequently the H_intent channel—and therefore the spatially aligned representation credited for the 10–15-step CT gains—is not generated from local raw sensors alone. Either the abstract/claims must be rewritten to state the communication and shared-localization assumptions explicitly, or the paper must add robustness experiments (noisy, delay","section":"Abstract; Sec. I; Sec. III-B (Eqs. 4, 8); Sec. IV-E"},{"comment":"Outdoor validation (Sec. IV-E, Fig. 1, Fig. 5) is valuable but does not close the perception loop claimed in the abstract. The virtual boundary is injected as artificial LiDAR points, and the evader’s relative state is taken from the shared Swarm-LIO estimate rather than an onboard detector. The paper acknowledges this as a simplification, yet still concludes that the framework is validated “using solely onboard sensing and computation.” A major revision should either (a) report at least one trial with an independent onboard evader detector (or a controlled noise model on the proxy) or (b) clearly demote the outdoor result to a coordination/control transfer test under privileged relative-state access, and adjust the abstract and conclusions accordingly. Without that, the hardware experiment does not fully support the perception-to-control narrative.","section":"Sec. IV-E; Abstract; Sec. V"},{"comment":"The efficiency advantage over OPEN (Table I: roughly 10–15 fewer capture steps) is attributed to “projection-level coupling” of intent and geometry (Sec. IV-C). OPEN receives privileged ground-truth obstacle states but processes them as disjoint tokens; PSTO receives dense LiDAR plus shared kinematic intent already projected into the same grid. The comparison therefore confounds representation design with information structure (privileged map vs. sensor grid; vector intent vs. spatially aligned heatmap). To make the causal claim load-bearing, the paper should either (i) give OPEN an equivalent dense obstacle encoding or (ii) add a controlled variant of PSTO that receives the same privileged obstacle map as OPEN while keeping the aligned intent channel. Absent that, the statement that PSTO is superior specifically because of spatial alignment remains only partially isolated, even though t","section":"Sec. IV-C; Table I; Fig. 4"}],"minor_comments":[{"comment":"Title and running text inconsistently space “AA V” / “AAV” (e.g., title “Multi-AA V”). Normalize to “AAV” throughout.","section":"Title; Abstract"},{"comment":"Fig. 2 system diagram is described but, from the manuscript text, the dual-stream backbone (dilated conv + SE on the heatmap stream) is only briefly justified as “empirically” better. A short quantitative comparison (shared CNN vs. dual-stream) in an appendix or table would strengthen Sec. III-C-1.","section":"Sec. III-C-1; Fig. 2"},{"comment":"Hyperparameters of PSTO (Vd×Hd = 120×6, λe, λt, Vevader, Vteammate, ddes, σcoord) and domain-randomization noise magnitudes are free design choices. Listing their values and any sensitivity study (even brief) would aid reproducibility.","section":"Sec. III-B; Sec. IV-A"},{"comment":"In Table I, at several high-speed/high-obstacle cells OPEN has slightly higher SR while PSTO has lower CT. The text correctly says “competitive success rates” and “superior capture efficiency,” but a one-sentence discussion of the SR–CT trade-off would prevent readers from over-reading the boldface pattern.","section":"Table I; Sec. IV-C"},{"comment":"Related work claims “to our knowledge, the first multi-AAV framework to learn a decentralized P–E policy directly from raw LiDAR data.” Soften or qualify given concurrent single-agent LiDAR RL and multi-agent state-based P–E lines already cited; the precise novelty is the unified PSTO for multi-agent aerial pursuit.","section":"Sec. II"},{"comment":"Eq. (1) writes the objective without the usual expectation over trajectories under the joint policy; a standard Dec-POMDP return notation would be clearer.","section":"Sec. III-A, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The technical core (PSTO + MAPPO curriculum + outdoor demo) is publishable after claim hygiene. The main risk is abstract-level overclaiming of “raw LiDAR only / onboard only” while the intent channel and outdoor evader state depend on shared kinematics—an assumption the body text discloses but the marketing language does not. I would not reject on novelty grounds; I would require the authors to align the abstract and conclusions with the actual information structure before acceptance. Scope is appropriate for a robotics systems/learning journal."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is that they built a single decentralized MAPPO policy that takes a two-channel egocentric spherical grid (raw LiDAR proximity + predictive intent heatmap) and outputs continuous velocity commands. It beats OPEN and other learning baselines that get ground-truth obstacles on capture time (roughly 10–15 fewer steps) while matching success rates, scales from 2-vs-1 to 4-vs-1 without retraining, and transfers zero-shot to outdoor quadrotors with onboard Mid-360 + NUC.\n\nWhat is actually new is PSTO itself: projecting dense LiDAR ranges and sparse multi-agent forecasts (LSTM for the evader, linear extrapolation for teammates) onto the same fixed-resolution spherical grid so the CNN can reason about geometry and intent together. The dual-stream backbone, progressive curriculum, and ablations (intent-blind, velocity-aware, separated-input) are clean and informative. Table I is thorough (150 rollouts across speed/obstacle grids); the outdoor pincer telemetry is real. Citations look honest; self-cites are to their prior single-agent navigation work and do not prop up the multi-agent claim.\n\nSoft spots are real but proportional. Teammate world-frame states are assumed shared over comms (Eq. 8) and outdoor evader state is a Swarm-LIO proxy rather than pure onboard detection (Sec. IV-E). So the H_intent channel is not generated from local raw sensors alone; if those links drop, the representation degrades. The paper is open about this as a hardware-aligned simplification, but it does soften the “only onboard sensing” slogan. No code release. Free parameters (grid size, decay factors, spacing, curriculum) are ordinary for this style of work and do not look like load-bearing fudge. Planar focus and shared-state assumption are stated limitations, not hidden ones.\n\nThis is for people building multi-UAV pursuit or end-to-end aerial MARL who care about representation design and sim-to-real. The central empirical claim holds under the stated assumptions. I would send it to peer review; the representation and outdoor result are worth referee time even if they must tighten the perception claims and release artifacts.","headline":"Solid engineering advance: PSTO + end-to-end LiDAR policy beats privileged SOTA on capture time, scales zero-shot, and flies outdoors; the pure-raw-sensor framing is only partial because intent still rides on shared kinematics.","tokens_in":15784,"tokens_out":559,"would_cite":true,"duration_ms":6198,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A single decentralized policy that reads raw LiDAR plus predicted intent can capture faster than systems that already know the obstacle map.","keywords":["multi-agent reinforcement learning","pursuit-evasion","aerial swarms","end-to-end control","LiDAR perception","decentralized policies","quadrotor"],"falsifier":"Run the identical outdoor 2-vs-1 trials with the communication link for teammate states disabled and with the evader tracked solely by each pursuer’s own LiDAR detector; if success rate or capture time then falls below the privileged-obstacle baselines, the end-to-end claim fails.","tokens_in":15700,"feed_emoji":"🚁","tokens_out":637,"duration_ms":6389,"temperature":0.7,"pith_summary":"Aerial pursuit teams usually cheat: they receive clean obstacle maps or hand-crafted geometric features before they decide how to fly. This paper argues that is unnecessary. It builds a fixed-size egocentric grid called Predictive Spatio-Temporal Observation (PSTO) that stacks raw LiDAR depth with short-horizon predictions of the evader and of teammates, then trains one shared deep policy that maps that grid straight to continuous velocity commands. In simulation the policy captures faster than strong baselines that still enjoy privileged obstacle information, keeps high success rates under faster evaders and denser clutter, and works for 2-, 3-, or 4-pursuer teams without retraining. Outdoor quadrotor flights with only onboard LiDAR and compute confirm the same behavior transfers. The practical claim is that spatial alignment of geometry and intent is enough to close the perception-to-control loop for cooperative aerial pursuit.","feed_headline":"Raw LiDAR plus predicted intent beats maps for drone pursuit","feed_subtitle":"One shared policy captures faster than systems that already know every obstacle, and flies outdoors.","key_machinery":"Predictive Spatio-Temporal Observation (PSTO): a fixed-resolution egocentric spherical grid whose two channels align dense LiDAR obstacle proximity with a sparse heatmap of predicted evader attraction and teammate repulsion; this single tensor is the sole perceptual input to a dual-stream convolutional actor.","core_discovery":"A decentralized multi-agent policy that maps the two-channel PSTO representation—raw LiDAR proximity plus a predictive intent heatmap—directly to continuous body-frame velocity commands achieves higher capture efficiency and competitive success rates versus state-of-the-art learning methods that receive privileged ground-truth obstacle states, while remaining a single network that scales zero-shot across different team sizes and transfers to outdoor quadrotor hardware.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Raw LiDAR and intent heatmaps beat maps for drone pursuit","One policy maps noisy LiDAR to multi-drone captures outdoors","PSTO grids let swarms intercept without privileged obstacle maps","Decentralized drones chase faster using predictive LiDAR grids","Shared policy scales aerial pursuit from raw scans to hardware"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Teammates must share their kinematic states over a communication link, and in the outdoor flights the evader’s relative position is taken from the shared state estimator rather than pure onboard detection; if either stream fails, the intent channel of PSTO collapses.","fun_headline_variants_meta":{"raw":{"variants":["Raw LiDAR and intent heatmaps beat maps for drone pursuit","One policy maps noisy LiDAR to multi-drone captures outdoors","PSTO grids let swarms intercept without privileged obstacle maps","Decentralized drones chase faster using predictive LiDAR grids","Shared policy scales aerial pursuit from raw scans to hardware"]},"model":"grok-4.5","effort":"low","cost_usd":0.005512,"raw_usage":{"total_tokens":1488,"prompt_tokens":758,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":55120000,"prompt_tokens_details":{"text_tokens":758,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":646,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":758,"tokens_out":84,"duration_ms":7006,"temperature":1.0,"reasoning_tokens":646,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T18:58:48.848126+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the identical outdoor 2-vs-1 trials with the communication link for teammate states disabled and with the evader tracked solely by each pursuer’s own LiDAR detector; if success rate or capture time then falls below the privileged-obstacle baselines, the end-to-end claim fails.","supporting_citations":[],"review_version":1}