{"id":"d9f2c5e1-ef76-418d-96c0-11b6d2327515","arxiv_id":"2608.01802","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A dual-altitude UAV system that treats cooperation as a leader-follower game improves vision-and-language navigation success rates in simulated cities.","lead":"This paper trains two drones to cooperate: a high-flying drone scouts from above and sends target coordinates to a low-flying drone that navigates to them. The team beats single-drone and earlier two-drone baselines in photorealistic city simulations.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed 'consistent' outperformance is not yet statistically grounded: per-tier test sets in Table 1 are roughly 10–15 targets each, so the 30.8pp OSR gap corresponds to a few target successes, and the 3-seed averages are reported without error bars or significance tests.","rationale":"Good-faith reading: the paper is a systems contribution with a clear pipeline, careful ablations (No-DINO, random retrieval, memory transfer), and an iteration curve showing complementary leader/follower updates; the promised code/data release is a plus. The central claim, however, is framed in absolute comparative terms ('consistently outperforms', up to +30.8pp) and the evidence is a small, author-constructed benchmark. By the paper's own Appendix J, the entire test pool is 84 instructions/106 targets across three scenes; per-tier cells are one or two dozen targets. At that scale, OSR gaps of 20–30pp can be caused by three to five targets, and 3 seeds do not fix this because the test episodes are presumably identical across seeds. The reader's identified weakest assumption (PPO expert as best-response proxy) is reasonable but not the most load-bearing for the empirical claim: even a suboptimal expert would produce a follower whose full-sim performance is exactly what Table 1 measures. The missing statistical evidence is what directly gates the strongest claim, so I would keep the CONDITIONAL verdict and add the release of per-episode results plus CIs as the gating condition. I partially agree with the reader: the benchmark weakness was mentioned in the rationale, but the formal weakest_assumption field points elsewhere.","tokens_in":23287,"tokens_out":7451,"duration_ms":70322,"concrete_test":"Release per-episode per-target outcomes (success/failure per GT target, per seed, per tier) for all CoNav-UAV and baseline rows in Table 1, and compute bootstrap 95% CIs for the OSR gap between CoNav-UAV and the best baseline within each scene-tier cell. If the CI for the airsim16 Hard gap (or the aggregate gap) includes 0, or the gap is driven by fewer than 5 target successes summed over seeds, the 'consistently outperforms' wording should be downgraded to a trend pending more episodes. A secondary check: state how many test instructions/targets fall in each tier and whether '3 seeded runs' vary the trained policy or only the evaluation randomness.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CoNav-UAV's central claim is empirical: it 'consistently outperforms' baselines with up to +30.8pp OSR, and the evidence for that claim is concentrated in Table 1. That table is built from only 252 instructions / 317 targets in total (Appendix J). Splitting each scene into Easy/Medium/Hard leaves roughly 10–15 test instructions/targets per cell on airsim16 (31 test instructions/45 targets) and fewer on airsim23/26. A 30.8pp gap on airsim16 Hard (37.2% vs 6.4% OSR) corresponds to about 5 target successes versus 1; the difference can easily flip if 2–3 episodes behave differently. Results are 'averaged over 3 seeded runs' with no CIs, and the paper does not state whether the three seeds share the same test episodes, so the effective sample for a gap is still the target count, not 3x. Without per-episode/per-target results, one cannot tell whether the headline gain is robust or a small-sample artifact. The PPO-expert transfer issue raised by the reader is real but secondary for the headline: the follower's behavior is evaluated in the full AirSim simulator, so a weak expert would cap performance rather than artificially inflate Table 1. The most load-bearing gap is the missing statistical support for the quantitative claim, not the proxy's optimality.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoNav-UAV, a dual-altitude UAV navigation system that models leader-follower cooperation as a general-sum Stackelberg game. A frozen-VLM leader performs BEV grounding and target dispatch using memory-augmented in-context learning, while a VLA follower executes coordinate-conditioned point-goal navigation trained by DAgger-style distillation from a PPO expert trained in a point-cloud geometric twin. Iterative Stackelberg Learning alternates these leader and follower updates. The system is evaluated in AirSim on three AerialVLN scenes (airsim16 as the learning scene, airsim23/26 for cross-scene transfer), reporting OSR improvements of up to 30.8 points on the learning scene and 9.0 points under transfer, with 2.6-3.7x less adaptation data than the baselines. The paper also releases code, data, and a benchmark annotation pipeline.","tokens_in":1661,"tokens_out":1728,"duration_ms":55121,"significance":"If the reported results are robust, the contribution is substantial: it is the first game-theoretic formulation of dual-altitude aerial navigation, it removes privileged information from the deployed system, and it provides a concrete alternating-optimization scheme matched to heterogeneous agent capabilities. The empirical protocol is also a strength: the authors compare against single- and dual-agent baselines, include held-out scenes, release code/data, and report ablations of memory transfer and components. However, the significance hinges almost entirely on the quantitative claim of consistent outperformance, and the evidence for that claim is currently not statistically grounded. The core contribution is therefore plausible but not yet demonstrated at the level the abstract claims.","major_comments":[{"comment":"The central empirical claim of consistent outperformance is not supported by the reported statistics. Per-tier test sets are very small: Table A7 lists only 45 test targets for airsim16, 29 for airsim23, and 32 for airsim26, split into three difficulty tiers, implying roughly 10-15 targets per cell. A 30.8pp OSR gap on airsim16 Hard (37.2% vs. 6.4%) corresponds to about 5 successful targets versus 1, so a handful of episode outcomes can flip the gap. Results are averaged over 3 seeded runs with no confidence intervals, no significance tests, and no per-seed or per-episode results; it is also not stated whether the three seeds use the same test episodes. The paper should report per-seed/per-episode numbers, confidence intervals or significance tests, and should qualify claims like consistently outperforms accordingly.","section":"Section 5.1, Table 1, Appendix J (Table A7)"},{"comment":"The method assumes that the PPO expert trained in the point-cloud geometric twin is a valid approximation of the follower's Stackelberg best response, but no evidence for this optimality gap is provided. The appendix describes the expert's architecture and training but reports no expert success rate, no comparison against classical planners or an oracle in the geometric twin, and no analysis of how close the expert is to optimal. Since the follower update is entirely a distillation of this expert, the headline claim of driving toward a Stackelberg equilibrium depends on this assumption. The paper should validate the expert's quality in the geometric twin and show that the DAgger-trained VLA actually approaches the expert's performance after distillation.","section":"Section 4.3 and Appendix D.2 (Eq. 2)"},{"comment":"The paper claims that Iterative Stackelberg Learning drives both agents toward a Stackelberg equilibrium, but no convergence argument or equilibrium-quality measure is provided. General-sum Stackelberg games can cycle under alternating best-response dynamics, and here the updates are heuristic: the leader's memory update is not an optimizer over the follower's response, and the follower's DAgger distillation is not a verified best response. The empirical Stackelberg rounds in Figure 2 show improvement in system OSR, but that does not demonstrate convergence to an equilibrium. I recommend either adding a formal convergence guarantee for the alternating scheme or explicitly reframing the contribution as an approximate, heuristic co-adaptation procedure rather than equilibrium computation.","section":"Sections 4.1 and 4.4"}],"minor_comments":[{"comment":"The abstract's consistently outperforms is stronger than Table 1 supports: in the airsim26 Hard tier, CoNav (Qwen3-VL-8B) achieves 10.1% OSR versus AerialVLA's 15.9%, and in airsim23 Hard the GPT-5.4-mini variant with memory (6.9%) is below its own no-memory variant (12.7%). The claims should be qualified to the strongest configuration or to five of six cross-scene tier combinations as stated in the body.","section":"Abstract and Section 5.2"},{"comment":"Table 2 reports that transferred memory improves OSR in all eight paired settings (scene-level averages), but Table A5 shows a per-tier regression for GPT-5.4-mini on airsim23 Hard (-5.8pp) and several near-zero gains. The scene-level averages hide meaningful per-tier variation; the text should acknowledge this and show per-tier results in the main body or footnote.","section":"Table 2 and Table A5"},{"comment":"The iteration notation (iter0, iter0.5, iter1, iter1.5, iter2) is used without a formal definition in the main text. Please state clearly which components are updated at each half-step and that each integer iteration corresponds to one DAgger follower update while half-integers are leader memory updates.","section":"Figure 2 and Tables A1-A3"},{"comment":"The claim of 2.6-3.7x less adaptation data compares trajectory counts, but the baselines' target-scene fine-tuning data includes language-annotated instruction-conditioned trajectories while CoNav-UAV's follower consumes only language-free coordinate trajectories; the comparison would be clearer if the data types and the role of the transferred leader memory were restated in the same sentence.","section":"Section 5.2"}],"recommendation":"major_revision","confidential_remarks":"The leader memory design is taken nearly verbatim from MemRL (Zhang et al. 2026), a self-cited work with overlapping authorship. The manuscript would be strengthened by a clearer statement of the incremental contribution of the leader-update component beyond direct reuse, and by an ablation that isolates the value-aware retrieval from the rest of the system. I do not see this as grounds for rejection, but the editor may wish to check novelty disclosure."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The real novelty here is the framing: modeling a frozen-VLM leader and a VLA follower as a general-sum Stackelberg game, with alternating gradient-free in-context learning for the leader and DAgger distillation for the follower. That is not in AeroDuo or the single-agent literature, and it is a practical, onboard-only recipe for cooperative aerial VLN. The authors also built a complete data-generation pipeline, ran careful ablations, and tracked per-iteration dynamics showing complementary gains from the two updates. Credit where earned: the system design is thoughtful and the implementation details are unusually complete.\n\nThe soft spots are mostly about evidence, not method. The central claim is empirical, and Table 1 is built on very small test sets. The whole benchmark has 252 instructions and 317 targets; the held-out split on airsim16 is 31 instructions/45 targets, and per-tier cells are roughly 10–15 targets. A 30.8pp OSR gap on airsim16 Hard corresponds to about five successes versus one. With three-seed averages and no confidence intervals or significance tests, \"consistently outperforms\" is stronger than the data supports. The ablation table also shows some per-tier regressions, e.g., GPT-5.4-mini on airsim23 Hard. That said, the stress-test note is right that the PPO-expert transfer issue is secondary: the follower is evaluated in the full AirSim simulator, so a weak expert would cap performance rather than inflate the headline numbers. The equilibrium wording is also looser than what is proven—ISL is a heuristic alternation, not a convergent algorithm—but the paper mostly frames this as approximate.\n\nThe citation pattern is fine; the MemRL overlap is disclosed and the borrowed memory mechanism is properly attributed. The promised code/data release plus per-episode results would address the main weakness.\n\nBottom line: this is a solid systems paper with a genuinely reusable formulation and a carefully documented pipeline. It deserves a serious referee and likely conditional acceptance, but only if the authors add error bars or significance testing, soften the universal-consistency claim, and make good on the release promise. I would bring it to a reading group mainly for the Stackelberg-alternation idea, and I would cite it in work on multi-agent VLN.","headline":"A genuinely new Stackelberg formulation for dual-altitude UAV cooperation with credible ablations, but the headline 'consistently outperforms' rests on far too few test targets to be statistically grounded.","tokens_in":24217,"tokens_out":1707,"would_cite":true,"duration_ms":17670,"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":"CoNav-UAV models dual-altitude drone cooperation as a leader-follower Stackelberg game and reports search success up to 30.8 points above single- and dual-agent baselines.","keywords":["vision-and-language navigation","unmanned aerial vehicles","Stackelberg game","multi-agent cooperation","imitation learning","DAgger","in-context learning","vision-language-action model"],"falsifier":"Deploy the distilled follower in the full AirSim simulator and compare its collision rate and minimum navigation error with the expert's performance in the point-cloud twin; if the follower's collision rate is substantially higher in the photorealistic world, the twin is not a faithful proxy for the best response. Alternatively, retrain the follower under an equally strong but different teacher, such as a sampling-based planner, and check whether the reported success margins survive; if they disappear, the specific expert rather than the Stackelberg alternation is the load-bearing component.","tokens_in":23088,"feed_emoji":"🚁","tokens_out":17278,"duration_ms":123196,"temperature":0.7,"pith_summary":"Target-oriented aerial navigation asks a drone to find an object from only a short text description of its appearance and surroundings, a task that demands both wide exploration and close-up, collision-free approach. CoNav-UAV argues these two demands cannot be reconciled in a single agent and instead splits them between two drones: a high-altitude leader that grounds candidate targets on bird's-eye imagery and dispatches coordinates, and a low-altitude follower that navigates to each dispatched target from a first-person view. The paper's central claim is that modeling this split as a general-sum Stackelberg game, in which the leader commits first and the follower best-responds, and optimizing it with alternating role-matched updates yields a system that consistently outperforms single- and dual-agent baselines, with success-rate gains of up to 30.8 points on the learning scene and 9.0 points under cross-scene transfer while using roughly three times less adaptation data. If true, cooperative dual-altitude search becomes a practical, onboard-only alternative to single-drone systems that rely on privileged information or human assistance.","feed_headline":"Leader-follower drone team beats baselines by up to 30.8 points","feed_subtitle":"A scout-and-navigator drone pair, trained as a Stackelberg game, finds targets with 3x less adaptation data.","key_machinery":"The central object is the dual-altitude Stackelberg navigation game, in which the leader's dispatch decisions define the follower's task and the follower's navigability defines the leader's payoff, so that the pair's target is a Stackelberg equilibrium: a leader commitment that is optimal in anticipation of the follower's best response. The mechanism that carries the argument is Iterative Stackelberg Learning, an alternating scheme with two role-matched optimizers. The leader, a frozen vision-language model, is updated without gradients: after each episode a reflector distills the rollout into a structured guideline, stored in an episodic memory whose $Q$-values are updated by an exponential moving average and retrieved by a composite similarity-and-utility score. The follower, a language-free vision-language-action model, is updated by DAgger-style distillation: a compact PPO expert trained in a point-cloud geometric twin generates collision-free waypoint sequences, which are replayed in the photorealistic simulator to render synchronized first-person images, and the follower is fine-tuned on those observation-action pairs. The alternation is deliberate, keeping the training distribution matched to deployment as the leader adapts to the follower's capability and the follower realigns with the leader's evolving dispatch behavior.","core_discovery":"The paper's central discovery, stated on its own terms, is that dual-altitude UAV cooperation for target-oriented vision-and-language navigation can be cast as a game in which the high-altitude UAV leads by committing to target dispatches and the low-altitude UAV follows by navigating to the dispatched coordinates, with the leader's reward tied to scene-level coverage and grounding precision and the follower's to efficient, collision-free arrival. It then claims that Iterative Stackelberg Learning, which alternates a gradient-free, memory-based in-context update for the frozen-VLM leader with DAgger-style expert distillation for the follower, drives the pair toward a Stackelberg equilibrium and delivers the reported results: success rates up to 40.6% on the easy tier of the learning scene, margins of up to 30.8 points over the strongest baseline in a tier, and the highest success rate in five of the six cross-scene tier combinations at a fraction of the baselines' adaptation cost. The paper also reports that the leader's learned guidelines transfer across scenes and across vision-language backbones, and that the two updates contribute complementary gains, including a transient regression when the leader's dispatch behavior moves ahead of the follower's current capability.","pith_inferences":["A decisive test of where the gains come from: if the leader's learned guidelines change materially when the follower's expert is replaced by a stronger or optimal teacher, then the game-theoretic alternation is amplifying follower quality rather than contributing independent reasoning value.","The round-based alternation with one DAgger epoch per round is a design choice; overlapped or more frequent follower updates might smooth the transient regression the paper observes, a testable modification of the algorithm.","The finding that weaker vision-language backbones saturate after the first guideline injection suggests the in-context ceiling is the backbone's spatial grounding; fine-tuning the leader for aerial grounding, which the authors list as future work, would reveal whether memory quality or model capability is the binding constraint.","Because the point-cloud twin is needed only for data generation, a single high-altitude survey pass could in principle build the twin and bootstrap follower training in a new scene, turning the leader's exploration into an enabler of its own teammate's training; the paper notes this direction but does not demonstrate it."],"forward_implications":["Cooperation can be trained with onboard visual and linguistic inputs alone: no ground-truth target coordinates, depth maps, or point clouds are needed at deployment.","Adapting to a new scene is cheap because the follower is language-free: roughly 2,000 coordinate-conditioned trajectories suffice, 2.6-3.7 times fewer than baselines that require new language annotations on the target scene.","The leader's memory-based guidelines encode scene-generic exploration and dispatch strategies that transfer across scenes and across vision-language backbones, so no per-scene leader retraining is needed.","A leader-only update can temporarily lower system performance until the follower catches up, implying that the reported gains come from the alternation itself rather than from either agent updated alone.","The same formulation extends to multiple followers, since the leader's dispatch queue generalizes as long as each follower best-responds to its dispatched target."],"supporting_citations":[{"why":"Supplies the three high-fidelity urban scenes from the AerialVLN benchmark that all experiments and comparisons run on.","marker":"(Liu et al. 2023)"},{"why":"AeroDuo, the only prior dual-altitude method and the strongest dual-agent baseline; the approach it must beat and whose privileged-information reliance it removes.","marker":"(Wu et al. 2025)"},{"why":"DAgger, the imitation-learning reduction whose alternating expert-distillation scheme the follower update is built on.","marker":"(Ross, Gordon, and Bagnell 2011)"},{"why":"PPO, the algorithm that trains the compact expert in the point-cloud geometric twin that supervises the follower.","marker":"(Schulman et al. 2017)"},{"why":"MemRL, the episodic-memory framework whose value-aware retrieval the leader's gradient-free in-context update adopts.","marker":"(Zhang et al. 2026)"},{"why":"The reduction of Stackelberg-equilibrium computation to alternating leader-follower updates that justifies Iterative Stackelberg Learning's structure.","marker":"(Gerstgrasser and Parkes 2023)"},{"why":"Qwen2.5-VL, the vision-language model the follower is initialized from and fine-tuned with LoRA.","marker":"(Bai et al. 2025)"},{"why":"OpenFly, the simulation platform the photorealistic environments are rendered on and a VLA baseline.","marker":"(Gao et al. 2025)"},{"why":"The leader-commitment game concept that the dual-altitude formulation is built upon.","marker":"(von Stackelberg 2011)"}],"fun_headline_variants":["Drone leader-follower pair lifts VLN success up to 30.8 points","CoNav-UAV: Stackelberg-trained drones need 3x less adaptation data","Dual-altitude drone team beats VLN baselines via Stackelberg learning","High-low drone duo improves target search by up to 30.8 points"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the compact PPO expert trained in the simplified point-cloud copy of the scene teaches the follower behaviors that still work in the full photorealistic simulator, including collision avoidance; if that proxy is not a faithful best response, the DAgger distillation and the reported gains rest on a weak teacher.","fun_headline_variants_meta":{"raw":{"variants":["Drone leader-follower pair lifts VLN success up to 30.8 points","CoNav-UAV: Stackelberg-trained drones need 3x less adaptation data","Dual-altitude drone team beats VLN baselines via Stackelberg learning","High-low drone duo improves target search by up to 30.8 points"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000245,"raw_usage":{"total_tokens":1959,"prompt_tokens":1161,"completion_tokens":798,"prompt_tokens_details":{"cached_tokens":1024},"prompt_cache_hit_tokens":1024,"prompt_cache_miss_tokens":137,"completion_tokens_details":{"reasoning_tokens":710}},"tokens_in":137,"tokens_out":798,"duration_ms":9749,"temperature":1.0,"reasoning_tokens":710,"cache_read_input_tokens":1024,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:04:21.009816+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the distilled follower in the full AirSim simulator and compare its collision rate and minimum navigation error with the expert's performance in the point-cloud twin; if the follower's collision rate is substantially higher in the photorealistic world, the twin is not a faithful proxy for the best response. Alternatively, retrain the follower under an equally strong but different teacher, such as a sampling-based planner, and check whether the reported success margins survive; if they disappear, the specific expert rather than the Stackelberg alternation is the load-bearing component.","supporting_citations":[{"cited_title":"In ConferenceonEmpiricalMethodsinNaturalLanguagePro- cessing (EMNLP)","cited_arxiv_id":null,"evidence_quote":"DAgger, the imitation-learning reduction whose alternating expert-distillation scheme the follower update is built on."}],"review_version":1}