{"id":"5d0d2354-3186-48f0-b8e6-6a4746b58052","arxiv_id":"2411.18009","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A monocular, edge-deployable reinforcement learning system lets fixed-wing UAVs avoid unknown obstacles at cruise speed, with modest success-rate gains over PPO baselines in simulation.","lead":"This paper trains a fixed-wing drone to dodge obstacles using only a single camera, flying at over 30 meters per second in simulation. It combines a depth-estimation network with a reinforcement learning controller and adds an adaptive exploration trick to improve success rates.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No end-to-end latency or depth-accuracy numbers are reported, so the central '30 m/s real-time avoidance' claim is unsupported.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the system assumes a single RGB frame yields a depth map accurate and timely enough at 30 m/s, but the paper does not evaluate depth error or latency. I agree this is the most critical point because the entire contribution is the deployable monocular system; if the perception loop cannot run within the required time budget, the '30 m/s' claim collapses regardless of the RL algorithm. Other issues exist, such as the undefined state-distribution ratio in Eq. 27 and the lack of error bars in Table II, but those affect the theoretical framing and statistical strength of comparisons rather than the fundamental feasibility of the system. The conclusion's own admission about sudden small obstacles further underscores that the depth module is the weak link. Since the paper is a simulation study with public code and the reader's conditional verdict already accounts for this missing evidence, I recommend no change to the verdict. The concrete test above would settle the concern by directly measuring the deployed pipeline's latency and depth quality in the same HIL setting.","tokens_in":14920,"tokens_out":2956,"duration_ms":27418,"concrete_test":"Instrument the HIL setup to log timestamps of image capture, depth inference, and action output on the OrangePi 5B for at least 1000 episodes, and report mean and p99 end-to-end latency. Additionally, compare inferred depth against AirSim ground-truth depth on 1000 held-out frames that include small obstacles; if p99 latency exceeds the simulator's control step or depth RMSE on near-field obstacles exceeds a threshold that would make a 3 m look-ahead impossible, the 30 m/s claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that single-frame monocular depth is both accurate and timely at 30 m/s. The paper never measures either. Section V-D describes HIL on an OrangePi 5B (RK3588s NPU) but reports no inference latency, frame rate, or control-loop period, and no comparison of inferred depth to ground truth. The depth model is ZoeDepth [45], a large transformer-based metric-depth network, fine-tuned but not replaced by a lightweight network; the lightweight StarNet-S1 only processes the already-generated depth map. At 30 m/s, a 100 ms perception delay means 3 m of flight; with the stated 60-timestep episodes and 224x224 depth, the pipeline may or may not keep up, but the paper provides no evidence. The conclusion concedes that sudden small obstacles are problematic, directly admitting the failure mode that matters at speed. Without latency and depth-error bounds, the empirical success rates in Table II cannot be attributed to a deployable system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a monocular vision-based obstacle avoidance system for fixed-wing UAVs. A single RGB frame is processed by a fine-tuned ZoeDepth network into a depth map, encoded by StarNet-S1, and fed together with target-relative distance and angle into a PPO-style policy. The paper introduces a hand-crafted reward composed of target, collision, distance-progress, and track-alignment terms, an 'inverse PPO' formulation that rewrites the advantage function from this reward, and an 'adaptive entropy' coefficient scaled by the success rate in a batch. Experiments use AirSim with JSBSim dynamics: ablations of the reward and entropy terms, comparisons with PPO, TRPO, A3C, DQN, and DDPG across three scenes (City, Line-cruising, Valley), and a hardware-in-the-loop deployment on an OrangePi 5B with an RK3588s NPU. The paper claims success rates up to 86%, smoother trajectories, and edge-device feasibility, and it releases source code.","tokens_in":15134,"tokens_out":6915,"duration_ms":59276,"significance":"If the empirical claims are sustained, the system would be a valuable proof-of-concept for low-cost, high-speed fixed-wing avoidance in unknown environments. The paper has real strengths: public code, a realistic AirSim/JSBSim training setup, multiple environment types, a hardware-in-the-loop implementation on an NPU, and ablations isolating the reward and entropy modifications. However, the central 'real-time at 30 m/s' claim is not accompanied by end-to-end latency, frame-rate, or depth-accuracy measurements, and the theoretical sections contain undefined constants, an unsupported importance-sampling expression, and an incomplete proof. As written, the contribution is a promising engineering system with preliminary evaluation rather than a fully validated algorithmic advance.","major_comments":[{"comment":"The real-time claim for cruise speed over 30 m/s is not supported by quantitative timing or depth-quality data. Section V-D describes the OrangePi 5B setup and shows trajectory comparisons, but it reports no inference latency for the ZoeDepth depth module, no frame rate, no control-loop period, and no quantitative comparison of inferred depth against ground truth. At 30 m/s, a 100 ms perception delay corresponds to roughly 3 m of travel, and the conclusion itself concedes that inferred depth may fail on sudden small obstacles. Please report end-to-end latency (depth inference plus encoding plus policy), depth error on representative frames, and a timing budget against the episode step rate actually achieved in the HIL experiments.","section":"V-D"},{"comment":"Theorem 4.1 and its proof are not mathematically coherent. The theorem statement uses constants eta, sigma, zeta, kappa, m, lambda, T, and k without defining them, and Eq. (22) contains an expression with V^m C_k / ||pi_theta(s,a)-pi_theta(s',a)||^m whose denominator can vanish; the limit in Eq. (23) refers to H^k_N, which is never defined. The proof of Theorem 4.1 is a list of inequalities with no derivation linking Lemma 4.1 to Eqs. (24)-(26), and the constant kappa in Eq. (26) depends on Exp(R), which is not defined. Either provide a rigorous statement with all constants specified and a valid proof, or remove the theorem and state clearly that the adaptive entropy term is a heuristic.","section":"IV-C"},{"comment":"The importance-sampling derivation in Eq. (18) and the clipped loss in Eq. (27) are not the standard PPO objective and are not justified. Eq. (18) introduces a state-distribution ratio pi_theta(s)/pi_theta'(s) alongside the usual action ratio, but no argument is given for why this ratio is tractable or unbiased; Eq. (27) then uses pi_theta/pi_theta' multiplied by pi_theta(s)/pi_theta'(s) with InA_theta, while standard PPO uses pi_theta(a|s)/pi_theta_old(a|s) with an advantage estimated under the old policy. The notation theta versus theta' is also inconsistent with Eq. (17), where theta is the pre-update parameter vector. Please clarify which distribution is old, which is new, and either derive the estimator rigorously or replace it with the standard clipped PPO objective.","section":"IV-B"},{"comment":"The definition and interpretation of the entropy term are internally inconsistent. Definition 4.1 states that higher H corresponds to greater utilization and lower H to greater exploration, but since H is a positive multiple of the Shannon entropy of the policy, higher H is the standard maximum-entropy direction of increasing exploration. Eq. (20) additionally scales H by Ms/Batch, the fraction of successful episodes in the batch, so the 'adaptive entropy' is a success-weighted entropy rather than an adjusted entropy coefficient; the text does not explain why a higher success rate should scale entropy. Please reconcile the direction of the effect and state precisely how the entropy coefficient changes with the success ratio.","section":"IV-C"},{"comment":"The policy comparison table reports success rates without confidence intervals, number of seeds, or significance tests. The differences are small (e.g., 86.0 vs 82.0 in City, 80.0 vs 76.0 in Line-cruising, 74.0 vs 69.0 in Valley), and with 100 trials per scenario a few random seeds could easily change the ordering. The header 'Success Rate (%, down-arrow)' also conflicts with the text, which implies higher is better. Please report mean and standard deviation over multiple seeds and a statistical test, or qualify the claim as a point estimate.","section":"V-C, Table II"}],"minor_comments":[{"comment":"InA is just the standard advantage A = Q - V; calling it 'inferring' is confusing. Please use standard notation or explain what is inferred.","section":"IV-B, Eq. (17)"},{"comment":"Use atan2 rather than arctan for alpha so that the angle covers the correct quadrant.","section":"III-A, Eq. (7)"},{"comment":"The action set contains seven yaw values from Eq. (9) plus the continue-last-action option, but Table I reports Action Dimension 8; the discrete action list should be stated explicitly.","section":"III-A, Eq. (9) and Table I"},{"comment":"The formula includes T, k, N, lambda, and m with no stated ranges; T is used both as horizon in Eq. (22) and in the proof's condition T >= 10 zeta kappa log 10 zeta. Please define all symbols.","section":"IV-C, Eq. (22)"},{"comment":"The claim that each test is repeated 100 times per scenario does not state how many seeds or random obstacle configurations were used; please specify.","section":"V-C"},{"comment":"The conclusion that the algorithm makes 'quick decisions in real-time' is asserted in the text but no timing measurement is reported; this should be either measured or removed.","section":"V-D"},{"comment":"Reference [40] cites a textbook for TRPO instead of the original paper (Schulman et al., 2015); please add the primary reference.","section":"References"},{"comment":"Typos such as 'Entrophy', 'pratical', 'ia a inferring', and 'outperforme' should be corrected.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is not ready in its current form because the claimed real-time performance rests on unmeasured latency and depth quality, and the theoretical section needs substantial revision. If the authors can add the missing measurements and either correct or remove the unsupported theorem, the empirical core could be salvageable; if not, rejection may be appropriate. The 'inverse PPO' terminology is also likely to mislead readers, since the method is essentially PPO with a hand-crafted reward and an entropy heuristic."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on DRL for fixed-wing avoidance. The actual system—single-frame monocular depth via fine-tuned ZoeDepth, a StarNet backbone, PPO with a shaped reward and adaptive entropy—is a legitimate engineering combination. The field needs concrete evidence on low-cost, vision-only avoidance at fixed-wing speeds, and this paper provides some: Table II shows the proposed method beating PPO, TRPO, A3C, DQN, and DDPG in success rate across three simulated scenes, with 100 trials per cell. That is a real empirical contribution, even if the margins over PPO (4–5 points) are modest and the paper does not report variance or significance. The HIL deployment on an OrangePi 5B, with the converted RKNN model, is useful evidence that the pipeline can run on edge hardware at all.\n\nNow the soft spots, in proportion. The theory section is the weakest part. The 'inverse PPO' objective, Eq. 30, reduces to standard PPO with a hand-designed reward and an entropy bonus whose coefficient is scaled by success rate. Theorem 4.1 and its proof are not usable as written: the statement mixes volume-of-ball terms into a claim about sample means, the proof jumps from an inequality to a conclusion without defined constants, and I would not send a student to it as a model of rigor. But that is a framing flaw, not a load-bearing flaw—the empirical claims do not rest on it.\n\nThe bigger issue is the central deployment claim. The system is said to avoid obstacles at 30 m/s, but the paper never reports end-to-end latency, frame rate, or depth accuracy. At that speed, a 100 ms perception delay is 3 m of travel, and the conclusion itself concedes that small, sudden obstacles are a problem. The HIL section describes the hardware but gives no hard numbers on inference time, control-loop period, or how the depth errors look against ground truth. So the 'real-time at 30 m/s' claim is plausible but unsupported. I would ask the authors to add those measurements.\n\nMinor: the code is public but not pinned to a commit, and fine-tuning details are incomplete. Reproducibility is partial but real.\n\nWho benefits? Practitioners building fixed-wing vision-based avoidance systems, and researchers wanting a baseline for monocular DRL navigation. The paper deserves a serious referee but needs revision—mainly tightening the theory or re-framing it as a reward-shaping contribution, and adding latency and depth-error numbers. I would send it out.","headline":"A solid systems paper with a real but modest empirical contribution, whose formal framing and deployment claims outrun the evidence.","tokens_in":15631,"tokens_out":631,"would_cite":true,"duration_ms":7634,"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":"The paper claims that a fixed-wing UAV can avoid unknown obstacles at cruise speeds above 30 m/s using only a single onboard camera, by combining one-frame monocular depth inference with an inverse PPO controller whose adaptive entropy…","keywords":["fixed-wing UAV","monocular depth estimation","deep reinforcement learning","PPO","obstacle avoidance","adaptive entropy","edge computing","hardware-in-the-loop simulation"],"falsifier":"A controlled HIL or field test that places a small obstacle below the depth model's reliable size range in the flight path with less lead time than the aircraft needs to start a turn at 30 m/s, and records whether the controller collides, would settle the central claim.","tokens_in":14697,"feed_emoji":"🛩️","tokens_out":7725,"duration_ms":64281,"temperature":0.7,"pith_summary":"This paper claims that a fixed-wing UAV can avoid unknown obstacles at cruise speeds above 30 m/s with nothing but an onboard RGB camera and a low-power edge processor. It replaces multi-frame depth stacks with a single-frame monocular depth estimate, and it modifies PPO into an \"inverse PPO\" whose reward function explicitly balances target approach with trajectory smoothness and whose entropy coefficient adapts during training to manage exploration. If the claim holds, fixed-wing obstacle avoidance no longer needs prior maps, LiDAR, or heavy computing, which would make low-altitude delivery and inspection safer and cheaper. The supporting evidence is software-in-the-loop and hardware-in-the-loop simulation, where the proposed method reports higher success rates than PPO, TRPO, A3C, DQN, and DDPG, and smoother paths than a sample-based planner in complex scenes.","feed_headline":"Depth from one frame steers fixed-wing UAVs past obstacles at 30 m/s","feed_subtitle":"One RGB image becomes a depth map; an inverse PPO controller beats five DRL baselines in simulation.","key_machinery":"The load-bearing object is \"inverse PPO\", a PPO variant that replaces the usual advantage estimate with an \"inferring advantage\" built from a hand-designed dense reward, and couples that loss to an adaptive entropy module. The reward has four terms: a target-arrival bonus, a collision penalty, a distance-progress term, and a path-alignment term that rewards staying close to the planned trajectory; the alignment term is what produces smooth fixed-wing turns. On the perception side, a fine-tuned monocular depth model turns one RGB frame into a depth map, and a lightweight NPU-friendly backbone encodes it into a 256-dimensional latent state that is concatenated with the goal distance and bearing. The adaptive entropy term scales exploration by the share of successful episodes in the current batch, which the paper argues counters PPO's tendency to over-rely on historical data.","core_discovery":"The central discovery claimed is that a complete monocular obstacle-avoidance loop, from one RGB frame to a feasible fixed-wing maneuver, can run on edge hardware at 30 m/s. The perception module is a fine-tuned single-frame depth estimator whose output is encoded by a lightweight backbone into a latent vector; the controller is a PPO variant whose advantage function is computed from a dense reward with target, collision, distance-progress, and trajectory-alignment terms, plus an adaptive entropy term. In the reported experiments the proposed method reaches 86%, 80%, and 74% success in city, line-cruising, and valley scenes, beating PPO (82%, 76%, 69%) and the other baselines, and the hardware-in-the-loop runs on a low-power NPU board show the policy producing smoother and shorter paths than a sample-based motion-primitives method in the two more complex scenes. The paper also states its own limit: inferred depth may not be reliable for sudden, small obstacles, and real-flight tests are not yet done.","pith_inferences":["A direct way to strengthen the result would be an ablation that feeds the controller ground-truth depth instead of inferred depth; the gap between those success rates would separate perception errors from policy errors, which the paper does not report.","The \"inverse\" in inverse PPO should not be read as inverse reinforcement learning: the reward is still hand-designed, and the advantage is inferred from that reward, not learned from demonstrations.","The adaptive entropy schedule is independent of the perception stack, so it could be transferred to other PPO-based navigation or control tasks where exploration collapses early.","Real-flight tests on a VTOL fixed-wing platform, which the paper names as future work, are the natural next check, because the HIL simulator cannot capture depth sensor noise and wind disturbances."],"forward_implications":["Fixed-wing UAVs can navigate unknown low-altitude environments without prior maps, LiDAR, or heavy sensors, relying on a single RGB camera and an edge NPU.","Single-frame depth inference avoids the latency and memory cost of multi-frame depth stacks, making real-time avoidance at cruise speed feasible on lightweight hardware.","The adaptive entropy schedule shortens training and reaches higher cumulative reward than fixed-entropy PPO, lowering the cost of policy learning.","The smoothness-aware reward reduces abrupt course changes, which lessens strain on fixed-wing control systems and energy consumption.","The same framework deploys on a low-power edge platform in hardware-in-the-loop tests, indicating the policy can run outside a desktop-class GPU."],"supporting_citations":[{"why":"Supplies the monocular depth estimation model that is fine-tuned to produce the single-frame depth map used as perception input.","marker":"[45]"},{"why":"Supplies the lightweight feature-extraction backbone that encodes the depth map into the latent state fed to the policy.","marker":"[46]"},{"why":"Defines the PPO objective, clipping mechanism, and importance sampling that the proposed inverse PPO modifies.","marker":"[25]"},{"why":"Provides the high-fidelity visual simulation environment used to generate RGB images and train the agent.","marker":"[48]"},{"why":"Provides the fixed-wing flight dynamics model used in training and hardware-in-the-loop simulation.","marker":"[49]"},{"why":"Serves as the sample-based motion-primitives baseline against which the proposed controller is compared in hardware-in-the-loop tests.","marker":"[50]"}],"fun_headline_variants":["Single-frame depth inference steers fixed-wing UAVs at 30 m/s","Inverse PPO plus one camera helps fixed-wing drones avoid obstacles","Edge-ready monocular system defeats unknown obstacles for fixed-wing UAVs","One RGB image powers obstacle-dodging fixed-wing UAVs at cruise speed","Monocular DRL navigates fixed-wing UAVs past obstacles in real time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that one RGB frame converted by the fine-tuned monocular depth model is accurate and timely enough for avoidance decisions at 30 m/s, including for sudden or small obstacles; the paper's conclusion concedes that inferred depth may fail precisely there.","fun_headline_variants_meta":{"raw":{"variants":["Single-frame depth inference steers fixed-wing UAVs at 30 m/s","Inverse PPO plus one camera helps fixed-wing drones avoid obstacles","Edge-ready monocular system defeats unknown obstacles for fixed-wing UAVs","One RGB image powers obstacle-dodging fixed-wing UAVs at cruise speed","Monocular DRL navigates fixed-wing UAVs past obstacles in real time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000693,"raw_usage":{"total_tokens":3175,"prompt_tokens":1022,"completion_tokens":2153,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":638,"completion_tokens_details":{"reasoning_tokens":2056}},"tokens_in":638,"tokens_out":2153,"duration_ms":14033,"temperature":1.0,"reasoning_tokens":2056,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:36:27.479102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled HIL or field test that places a small obstacle below the depth model's reliable size range in the flight path with less lead time than the aircraft needs to start a turn at 30 m/s, and records whether the controller collides, would settle the central claim.","supporting_citations":[{"cited_title":"Rewrite the stars","cited_arxiv_id":null,"evidence_quote":"Supplies the lightweight feature-extraction backbone that encodes the depth map into the latent state fed to the policy."},{"cited_title":"Airsim: High-fidelity visual and physical simulation for autonomous vehicles","cited_arxiv_id":null,"evidence_quote":"Provides the high-fidelity visual simulation environment used to generate RGB images and train the agent."},{"cited_title":"Jsbsim: An open source flight dynamics model in c++","cited_arxiv_id":null,"evidence_quote":"Provides the fixed-wing flight dynamics model used in training and hardware-in-the-loop simulation."},{"cited_title":"Motion primitives and 3d path planning for fast flight through a forest","cited_arxiv_id":null,"evidence_quote":"Serves as the sample-based motion-primitives baseline against which the proposed controller is compared in hardware-in-the-loop tests."}],"review_version":1}