{"id":"042e9686-f7a9-4ce7-9e33-32e392cda0f0","arxiv_id":"2501.07566","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"SafeSwarm applies MAPPO to multi-drone landing with obstacle avoidance, reporting 2.25 cm error in five real-world flights, but the claimed safety barrier algorithm is not described.","lead":"This paper presents SafeSwarm, a reinforcement learning system for landing a swarm of drones on moving pads while avoiding obstacles. The authors report 2.25 cm landing error in a small indoor test, but the core 'safe barrier net' algorithm is never defined.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'safe barrier net algorithm' is never specified; safety and collision-free performance are attributed to a standard shaped MAPPO reward, and the real-world test appears to use a single drone rather than a swarm.","rationale":"I agree with the reader's rejection, and the reader's weakest assumption is exactly the load-bearing concern: the absence of the safe barrier net. The paper promises a safe RL mechanism but never specifies it, so the reader cannot verify the main technical contribution. This is not a matter of disagreement with consensus; it is a missing, load-bearing component. The reward function is a standard shaped reward and cannot, by itself, substantiate a 'safe barrier net' or collision-free guarantee. The real-world evidence further narrows the claim to single-drone landing, leaving swarm collisions as an unvalidated simulation result. A single concrete check -- reimplementing with and without the claimed barrier net, and requiring a true multi-drone hardware trial -- would settle whether the barrier net contributes anything. Until then, the paper should not be accepted as a demonstration of safe swarm landing.","tokens_in":5699,"tokens_out":3723,"duration_ms":36313,"concrete_test":"Ask the authors to provide the barrier-net module's exact architecture, forward pass, and training/loss specification, or to release code. Then independently implement two versions of the system in gym-pybullet-drones with identical MAPPO and reward Eq. (1)-(6): one with the barrier-net safety filter active and one without. If success rate, collision count, and precision do not differ significantly across 30 or more runs with 3-4 drones and 3 obstacles, the barrier net is not responsible for the reported safety. Separately, require a hardware trial with at least three Crazyflies landing simultaneously; if only single-drone flights are produced, the 'collision-free landings' claim must be downgraded to single-agent landing.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is a 'safe barrier net algorithm' (abstract, Sec. I), but no such algorithm is defined. Section III.A only mentions using PPO and TD3 'with the Control Barrier Net function' without architecture, training objective, or safety constraint. The only implementation detail is the reward function in Eq. (1)-(6), which combines distance-to-target, collision, edge, and velocity penalties; these are heuristics, not a safety layer or barrier certificate. Consequently, the claim of 'collision-free landings' is not backed by any mechanism that could guarantee safety. The real-world evaluation also undercuts the swarm claim: Section IV.B describes one Crazyflie quadrotor, five flights, and Fig. 4 shows a single trajectory, so the swarm-level collision-free result is only simulated. With no barrier net, no multi-drone hardware data, and no independent baseline or code, the paper reduces to a MAPPO landing experiment, and the stated claims of safe decentralized swarm landing are unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SafeSwarm, a decentralized multi-agent reinforcement learning system intended to land swarms of Crazyflie drones on moving landing pads in cluttered environments. The authors introduce a reward function with penalties for distance, collisions, edge proximity, and velocity, and they report a simulation comparison with the MARLander baseline as well as real-world experiments. The abstract and introduction name a 'safe barrier net algorithm' and claim centimeter-level landing accuracy and collision-free landings. However, the manuscript never specifies the barrier-net algorithm, the real-world evaluation uses a single drone rather than a swarm, and no collision metric is reported for the simulation experiments.","tokens_in":5865,"tokens_out":3983,"duration_ms":41676,"significance":"If the claims were fully supported, this work would be a useful application of multi-agent reinforcement learning to safety-critical drone landing, with a concrete quantitative result of 2.25 cm mean landing error and a comparison against a prior MARLander baseline. The paper does report an actual physical experiment and gives numeric landing precision, which is a strength. However, the central novelty named in the title and abstract, the 'safe barrier net algorithm', is not described or evaluated. The swarm-level claim is not tested on hardware, the baseline is from the same group, and no variance or collision statistics are given. As written, the contribution reduces to a reward-shaped MAPPO landing experiment with simulated multi-agent results, so the significance of the stated claims is not established by the presented evidence.","major_comments":[{"comment":"The manuscript never specifies the 'safe barrier net algorithm' that is named in the Abstract and Introduction. Section III.A only states that PPO and TD3 were used 'with the Control Barrier Net function(CBF)', and Section III.C gives a shaped reward function in Eqs. (1)-(6) that contains distance, collision, edge, and velocity penalties. No barrier function, safety layer, action projection, formal guarantee, or training objective is defined. Since the safety and collision-free claims are the central novelty, this is a load-bearing omission. The authors should either provide the full algorithm and how it constrains actions, or revise the claims to those of a reward-shaped MAPPO policy.","section":"III.A and III.C, Eq. (1)"},{"comment":"The real-world evaluation described in Section IV.B uses one Crazyflie drone, five flights, and Fig. 4 shows a single trajectory, yet the title and abstract claim a swarm of drones with collision-free landings. Thus, swarm-level collision avoidance is not validated on hardware in any way. The 'collision-free landings' claim in the abstract is therefore unsupported by the reported experiments. The authors should either add multi-drone hardware trials with inter-agent collision measurements or explicitly limit the real-world claim to a single-drone landing with the swarm behavior demonstrated only in simulation.","section":"IV.B, Fig. 4"},{"comment":"The only baseline, MARLander [12], shares authors with this paper, so the comparison is not independent. In addition, the simulation results are reported as point estimates without standard deviations, confidence intervals, or the number of training seeds, despite the claim that the task was run 30 times. Success rate, precision, and time are not sufficient to support the comparative statements without variance information. The authors should add seed-level statistics and at least one independent baseline, such as a vanilla MAPPO without reward shaping or a known safety-layer method, before the relative safety and precision claims can be assessed.","section":"IV.A, Tables I and II"},{"comment":"No collision metric is reported in the simulation results. 'Collision-free' and 'success rate' are not the same; a landing attempt can be successful while still involving a collision, or a collision can be avoided at the cost of a failed landing. The manuscript should define how collisions with obstacles and between agents are detected in simulation and report the number or rate of collisions per episode. Without such a metric, the safety claim is a qualitative statement rather than a measured result.","section":"IV.A, Tables I and II"}],"minor_comments":[{"comment":"The terms αi and βi are called coefficients but are defined as reciprocal-distance factors; please clarify their role and specify the units and ranges of λ, ε, and γ.","section":"III.C, Eqs. (3)-(6)"},{"comment":"The sentence 'The loss of value function reached 0.01 at the and of training' contains a typo; 'and' should be 'end'.","section":"IV.A"},{"comment":"Figures 2 and 3 do not label their axes or give units, and the mean reward value of -2000 is not contextualized. Adding axis labels and seed variance would make the training curves interpretable.","section":"IV.A, Figs. 2 and 3"},{"comment":"The phrase 'Control Barrier Net function(CBF)' conflates the Control Barrier Function concept with the BarrierNet architecture; please cite and define the exact formulation used.","section":"III.A"},{"comment":"The statement 'accuracy of landing 80%' is ambiguous in relation to the reported 2.25 cm mean error; please define the success threshold used to compute the accuracy.","section":"IV.B"}],"recommendation":"reject","confidential_remarks":"The central safety mechanism named in the title and abstract is not present in the manuscript, the only baseline is the authors' own MARLander work, and the hardware experiment does not involve a swarm. These are load-bearing gaps that cannot be fixed by local edits. A substantially expanded version with a fully specified barrier-net algorithm, independent baselines, seed-level statistics, collision metrics, and multi-drone hardware validation would be needed for a fresh review."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: the paper reports a real MAPPO-based landing system for a single quadrotor in an indoor cluttered space, with decent cm-level accuracy on five flights, but the advertised “safe barrier net algorithm” is never defined anywhere in the body, and the swarm-level collision-free claim is only simulated. The reader’s REJECT verdict is essentially right.\n\nWhat is actually new: the paper applies MAPPO to the specific task of landing in a static-obstacle environment with a moving pad, and it shows a real Crazyflie experiment with Vicon tracking. The 2.25 cm mean landing error on five flights is plausible and a modest empirical step. The reward shaping in Eq. (1)–(6) is explicitly written out, and the comparison to MARLander, though same-group, gives some head-to-head numbers. That is more than a pure idea paper.\n\nThe soft spots are load-bearing. The “safe barrier net” appears only in the abstract and a single aside in Sec. III.A; there is no network architecture, training objective, safety constraint, or ablation. The safety guarantee therefore reduces to the collision-penalty term in the reward. That is a heuristic, not a barrier certificate. No code or data are provided, so the reader cannot check whether anything like a barrier net exists. The real-world experiment uses one drone, not a swarm, so the title’s promise of safe swarm landing is unsupported by hardware. The simulation results show 30 runs with no variance, no seeds, and only a single baseline (the authors’ own MARLander). The claim of “collision-free landings” is not measured at the swarm level.\n\nTo be fair, the paper does not appear to fabricate numbers; the landing errors and times are consistent with typical Crazyflie performance. But the gap between the claims and the evidence is central, not cosmetic. As written, the contribution reduces to a MAPPO landing experiment with a shaped reward.\n\nWho this is for: someone building an RL drone-landing benchmark might look at the experimental setup, but the missing algorithm makes it unusable as a reference. I would not send this to peer review in its current form. The authors should either properly specify the barrier net (and show it in ablation) or reframe the paper as a benchmark study with thorough statistics and an independent baseline. If they do that, a resubmission could be worth a look.","headline":"Reported MAPPO landing results are real but the 'safe barrier net' is never specified, and the swarm claim is simulated; the paper is not ready for review.","tokens_in":6435,"tokens_out":4348,"would_cite":false,"duration_ms":40446,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Decentralized reinforcement learning can land a swarm of small drones on moving platforms in cluttered scenes, with 2.25 cm mean error and no collisions in the reported tests.","keywords":["safe reinforcement learning","swarm of drones","autonomous landing","multi-agent reinforcement learning","collision avoidance","moving landing pad","decentralized control","MAPPO"],"falsifier":"Run the reported landing task with at least two drones flying simultaneous trajectories through the same obstacles in the real indoor setup and record all inter-agent distances; any mid-air collision or a landing error well above 2.25 cm would contradict the collision-free, centimeter-accuracy claim. A second check is to inspect the released code for a barrier layer distinct from the reward function; if no such layer exists, the claimed safety mechanism is not implemented.","tokens_in":5489,"feed_emoji":"🚁","tokens_out":10797,"duration_ms":93057,"temperature":0.7,"pith_summary":"This paper proposes SafeSwarm, a decentralized reinforcement-learning controller that trains small quadrotors to land on moving landing pads while avoiding static obstacles. The authors report a mean landing error of 2.25 cm with a mean landing time of 17 s in indoor flights on a micro quadrotor, with 80% landing accuracy. In simulation the same policy handles a swarm of agents and beats a prior multi-agent landing baseline on precision under both static and moving platforms (95% vs 91.67% success at 2.25 vs 2.26 cm; 80% vs 75% at 3.04 vs 3.93 cm). The intended contribution is that safe landing behavior can come from a single decentralized learned policy rather than from controllers that need each agent to know the other agents' dynamics.","feed_headline":"Drone swarm RL lands on moving pads to 2.25 cm","feed_subtitle":"Decentralized policy trained in simulation lands real drones while dodging obstacles, with centimeter-level error.","key_machinery":"The behavior is carried by the reward function of Eq. (1), with four terms: an encouragement term that pulls the drone toward the target pad, a constant penalty for collision, an edge penalty for being under the pad, and a velocity penalty that couples the drone's speed to its distance from the nearest obstacle and target (Eqs. 2-6). On top of this, the paper names a 'safe barrier net algorithm' as the safety mechanism that produces collision-free landings, although the text gives no formal definition of that barrier. The policy itself is MAPPO with an MLP network, run in a decentralized way so each agent acts on its own observations.","core_discovery":"The central claim is that a MAPPO-based decentralized policy, shaped by the reward in Eq. (1), can make a swarm of drones with different dynamics land on moving platforms in cluttered scenes without hitting obstacles or each other. The paper reports simulation success rates of 95% on a static platform and 80% on a moving platform, with higher precision than the compared baseline, and a real-world experiment in which a micro quadrotor landed with 2.25 cm mean error and 80% success in 17 s. The authors attribute these results to their 'safe barrier net algorithm' combined with a reward that encourages approaching the pad, penalizes collisions and under-pad positions, and slows the drone near obstacles and targets.","pith_inferences":["The hardware evidence in the paper is a single-drone flight, so the swarm-level collision-free claim is supported only by simulation; a two-or-more-drone hardware test would be the direct check of that claim.","Because the safety behavior is attributed to reward heuristics rather than to a formally verified barrier, the 'safe' label would be stronger if a control barrier function with provable invariance were added to the policy.","If the barrier-net specification is supplied in a later version, the same framework could be adapted to delivery-drone operations or indoor inspection fleets where guarantees about inter-agent clearance matter.","The velocity penalty mechanism implies a testable prediction: landing accuracy should degrade gradually as localization noise increases, since the policy's obstacle avoidance depends on the quality of relative-position estimates."],"forward_implications":["If the reported result transfers to multi-drone flights, a single decentralized policy can coordinate heterogeneous drone teams without inter-agent dynamics models, removing a major scalability bottleneck of model predictive control for landing.","The 2.25 cm landing error is the same order as motion-capture-based single-drone landers, so obstacle avoidance does not obviously sacrifice touchdown precision.","Because the policy is trained in simulation and works on hardware in the reported tests, the approach is a candidate for sim-to-real deployment on resource-limited aerial robots.","Moving-platform tests suggest the method keeps a precision edge over the prior swarm-landing baseline even when the target moves, at the cost of a longer mean landing time (21 s vs 17 s).","The reward shaping recipe of distance encouragement, collision penalty, under-pad boundary, and proximity-scaled velocity is a reusable pattern for other cluttered-environment landing tasks."],"supporting_citations":[{"why":"Supplies the prior swarm-landing planning method that SafeSwarm is compared against in the static- and moving-platform tables.","marker":"[12]"},{"why":"Defines differentiable control barrier functions for safe robot learning, the family of methods the named 'safe barrier net' draws on.","marker":"[2]"},{"why":"Introduces an attention barrier net for safe and scalable robot learning, framing the safety mechanism class the paper claims to implement.","marker":"[1]"},{"why":"Provides the single-agent PPO decentralized-landing baseline that the swarm approach extends.","marker":"[19]"},{"why":"Supports the simulation-to-reality transfer argument that a policy trained in simulation can work on physical drones.","marker":"[15]"}],"fun_headline_variants":["Safe RL lands drone swarm on moving pads to 2.25 cm","Drone swarm lands on moving pads with 2.25 cm error","Safe decentralized RL lands drone swarm on moving pads","Safe barrier net enables cm-accurate swarm landing on moving pads","MAPPO swarm lands on moving pads with 2.25 cm accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on the existence and effectiveness of the named 'safe barrier net algorithm', but the paper never specifies how it works; if that net is not actually present, the safety novelty reduces to a standard MAPPO training run with reward shaping.","fun_headline_variants_meta":{"raw":{"variants":["Safe RL lands drone swarm on moving pads to 2.25 cm","Drone swarm lands on moving pads with 2.25 cm error","Safe decentralized RL lands drone swarm on moving pads","Safe barrier net enables cm-accurate swarm landing on moving pads","MAPPO swarm lands on moving pads with 2.25 cm accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.002328,"raw_usage":{"total_tokens":8920,"prompt_tokens":832,"completion_tokens":8088,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":448,"completion_tokens_details":{"reasoning_tokens":7998}},"tokens_in":448,"tokens_out":8088,"duration_ms":48226,"temperature":1.0,"reasoning_tokens":7998,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:37:51.679176+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the reported landing task with at least two drones flying simultaneous trajectories through the same obstacles in the real indoor setup and record all inter-agent distances; any mid-air collision or a landing error well above 2.25 cm would contradict the collision-free, centimeter-accuracy claim. A second check is to inspect the released code for a barrier layer distinct from the reward function; if no such layer exists, the claimed safety mechanism is not implemented.","supporting_citations":[{"cited_title":"MARLander: A Local Path Planning for Drone Swarms using Multiagent Deep Reinforcement Learning","cited_arxiv_id":"2406.04159","evidence_quote":"Supplies the prior swarm-landing planning method that SafeSwarm is compared against in the static- and moving-platform tables."},{"cited_title":"Morpholander: Reinforcement learning based landing of a group of drones on the adaptive morphogenetic uav,","cited_arxiv_id":null,"evidence_quote":"Provides the single-agent PPO decentralized-landing baseline that the swarm approach extends."},{"cited_title":"Deep drone racing: From simulation to reality with domain randomization,","cited_arxiv_id":null,"evidence_quote":"Supports the simulation-to-reality transfer argument that a policy trained in simulation can work on physical drones."}],"review_version":1}