{"id":"899c64af-2318-4e27-84a5-9ac48ebba838","arxiv_id":"2509.11240","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CORB-Planner uses safe flight corridors as low-dimensional observations for an RL policy that generates B-spline control points, enabling real-time cross-platform UAV planning after about ten minutes of training.","lead":"A team built a drone flight planner that uses simplified corridor shapes around a path as the input to a reinforcement learning agent, letting it generate smooth obstacle-avoiding trajectories in real time. The planner trained in about ten minutes in simulation and then steered four different drone types at up to 8.2 meters per second in cluttered real-world settings, which could make learning-based navigation cheaper and more portable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The sim-to-real transfer rests on an unvalidated perfect-tracking assumption; if real tracking error approaches SFC clearance, the promised cross-platform high-speed results are not supported.","rationale":"I read the paper in good faith. The architecture is coherent: A* plus SFC gives a low-dimensional observation, the RL policy outputs acceleration control points, and a low-level controller tracks the resulting B-spline. The ten-minute training claim is plausible given the simple MLP and multi-threaded sampling, though unverified. The main risk is the perfect-tracking assumption: the training environment is purely kinematic, and the paper's own limitation statement admits it. Real quadrotors track with lag and error; in dense clutter at 8.2 m/s, a 0.2-0.5 m error could exceed the SFC clearance. The paper provides no quantitative support that the policy is robust to this. The real-world demos are encouraging but anecdotal—no tracking error data, no repeated trials, no statistics. Other issues (the 8.2 m/s vs 2 m/s discrepancy, missing error bars, no code release) are addressable and secondary to the tracking assumption. Therefore the verdict remains conditional: the central claim is plausible but not established until the tracking assumption is validated. I agree with the reader's identification of the weakest assumption, and no change to the reader's verdict is needed.","tokens_in":14154,"tokens_out":7518,"duration_ms":91759,"concrete_test":"Reproduce the dense-wall and forest simulation benchmarks (Table I) with the trained vmax=7 and vmax=10 CORB_S policies, but add a first-order tracking lag (e.g., time constant 0.1 s) to the executed B-spline before checking SFC violations. If success rates drop by more than ~20 percentage points relative to the no-lag baseline, the perfect-tracking assumption is load-bearing and the sim-to-real transfer claim is unverified; if rates remain comparable, the concern is mitigated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-B explicitly states the training environment 'does not incorporate physical simulations' and that 'the tracking of B-spline trajectories is assumed to be perfectly precise'; only 'random noise into the planning process' is added, not dynamics noise. The RL policy is therefore optimized under kinematic B-spline transitions (Eqs. 2, 8), with no model of closed-loop tracking error. The cross-platform claim (Section V-D) and the 8.2 m/s forest flight (Section V-E) require the onboard PID+PX4 controller to track the generated B-spline closely enough that the vehicle remains inside the SFC. The paper provides no measurement of tracking error, no SFC clearance margin analysis, and no robustness evaluation against tracking lag. If tracking error is comparable to the corridor's free space, especially in dense clutter at high speed, the policy's decisions become infeasible and the central claim fails. This is the single most load-bearing concern because all real-world results depend on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents CORB-Planner, a real-time RL-based B-spline trajectory planner for UAVs. The observation is a compact safe flight corridor (SFC) derived from an A* reference polyline, together with the current B-spline control points and time. An RL policy outputs acceleration control points that are transformed into B-spline control points; a reward function penalizes SFC violations, rewards progress along the SFC, and applies a jerk discount. Training uses the authors' SDCQ algorithm in a physics-free easy-to-hard curriculum with exploitation-decoupled multi-thread sampling, claiming convergence in about ten minutes. Experiments compare against EGO-planner in simulation, ablate algorithm/training choices, and report real-world flights on four UAV platforms, including an 8.2 m/s forest flight and a 275 g quadrotor with a Cortex-A53 board.","tokens_in":14434,"tokens_out":6509,"duration_ms":74744,"significance":"If the central claims hold, the paper demonstrates a practical route to cross-platform, low-compute RL trajectory planning: a single simulation-trained policy transfers to quadrotors, hexarotors, LiDAR, and vision sensing, and runs at 25 Hz on a lightweight board. The SFC-as-observation idea is a clean and credible mechanism for reducing overfitting to platform-specific sensing and dynamics. The paper provides external baselines (EGO-planner, DPMPC-planner), ablations against PPO/SAC, and physical experiments on multiple platforms, which are significant assets. However, the absence of code/data and the unvalidated perfect-tracking assumption weaken the reproducibility and quantitative support for the cross-platform claim.","major_comments":[{"comment":"The training environment 'does not incorporate physical simulations' and 'the tracking of B-spline trajectories is assumed to be perfectly precise'; only random noise is added to the planning process. The real-world transfer claims (Section V-D and the 8.2 m/s forest flight in Section V-E) rest on the onboard PID+PX4 controller tracking the generated B-spline closely enough that the vehicle remains inside the SFC. No tracking-error measurement, SFC clearance margin, or sensitivity analysis is reported. If the tracking error is comparable to the corridor's free space, the policy decisions can become infeasible. Please report measured position tracking error versus SFC widths/clearances for each real-world platform and speed, and/or add a tracking-error model or dynamics noise in training and re-evaluate. This is load-bearing for the central cross-platform and high-speed claims.","section":"Section IV-B, with Section V-D/E"},{"comment":"As printed, line 7 'Add u_k to u*' is inside the for loop, so the final waypoint u_k is appended at every iteration, producing a corrupted polyline. The polyline directly determines the SFC used in the observation and reward, so this is not merely cosmetic. Please correct the pseudocode by moving line 7 after the loop and verify the loop range, which currently starts at i=2 and appears to skip u_1.","section":"Algorithm 1 in Section III-A"},{"comment":"The discrete action index is defined as α_d^t ∈ [1,M]^3 and the mapping is α_t = (2α_d+1)/M−1. For α_d=M and M=60, this gives α_t ≈ 1.0167, outside the claimed cubic action space [−1,1]^3; for α_d=1 it gives −0.95. If implemented literally, the action bounds are inconsistent with Section III-B. Please clarify the index convention (e.g., α_d ∈ [0,M−1] with mapping (2α_d+1)/M−1, or use 2(α_d−1)/(M−1)−1 to exactly cover [−1,1]).","section":"Section IV-A, Eq. (18)"}],"minor_comments":[{"comment":"The text introduces coefficients 'k_p, k_v, and k_s', but Eq. (9) uses k_f. Please unify the notation.","section":"Section III-C"},{"comment":"The orthogonal vector n_i contains subscript typos: 'u^*_{i,y-1}' and 'u_{i,y}' should likely be 'u^*_{i-1,y}' and 'u^*_{i,y}'. Please fix.","section":"Section III-A, Eq. (4)"},{"comment":"The vision-based quadrotor is described with a 'RealSense D430' here, while the Introduction and Section V-D say 'RealSense D435i'. Please make the sensor name consistent.","section":"Section V-A"},{"comment":"The text refers to 'Fig. I' but should refer to 'Fig. 1'.","section":"Section V-A"},{"comment":"The cited title is 'Soft Decomposed Policy-Critic', but the paper describes 'Soft Decomposed-Critic Q (SDCQ)'. Please align the reference title with the algorithm name.","section":"Reference [14]"},{"comment":"No code or data release is mentioned. Given the ten-minute training claim and the cross-platform transfer result, a reproducibility statement or supplementary material (e.g., trained policy parameters, training configuration, and a sample real-world dataset) would substantially strengthen the paper.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper relies heavily on the authors' own SDCQ algorithm [14] and the same group's Storm [22]; the external validation is against EGO-planner and DPMPC-planner. The lack of code/data release is a reproducibility concern, and the perfect-tracking assumption needs quantitative support. The central idea is promising and the empirical results are suggestive, but the load-bearing tracking-error gap and the Algorithm 1 pseudocode error preclude acceptance in the current form."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about CORB-Planner. First, it is a useful engineering contribution: the SFC-as-observation representation is compact and sensible, the cylindrical action normalization is a neat fix for non-uniform dynamics, and the ten-minute training claim is backed by ablations and real flights on four different UAVs. Second, the paper's biggest weak spot is exactly what the stress test flags: training assumes perfect B-spline tracking and only adds noise to the planning process, not to the dynamics. Without tracking-error measurements or clearance-margin analysis, the cross-platform transfer claim rests on the hope that any low-level controller tracks well enough. The real flights show it works for these platforms, but the paper does not quantify how close the operating margin is.\n\nThe genuinely new piece is the specific integration: using a 66-dimensional safe-flight-corridor observation to feed a value-based discrete RL policy that emits B-spline control points, plus the exploitation-decoupled easy-to-hard curriculum. The paper is honest about leaning on the authors' own SDCQ algorithm, and the validation against EGO-planner, DPMPC-planner, PPO, and SAC is appropriate. The physical experiments, including the 275 g quadrotor with a Cortex-A53 board, give real evidence that the method runs on modest compute.\n\nThe soft spots are real but mostly addressable. The abstract's \"up to 8.2 m/s\" refers to the high-performance quadrotor in the forest, not to the lightweight board, which only flew at 2 m/s; that is an overstatement that should be fixed. Eq. (7) appears to lack absolute values in the max, a typo that matters for negative actions. No code or data is released, and Table I has no error bars. The more substantive issue is the perfect-tracking assumption: if tracking error approaches corridor clearance in dense clutter, the policy's decisions become infeasible. I would want to see tracking-error plots for the real flights, or at least a discussion of how the SFC margins absorb them.\n\nOverall, the central idea holds up and the paper deserves a serious referee. It is not a paradigm shift, but it is a solid within-subfield result. I would recommend accepting it for peer review, with the caveat that the authors should release artifacts, correct the speed claims, and either add tracking-error data or weaken the cross-platform generality claim to match what is demonstrated.","headline":"A credible RL-based B-spline planner with real cross-platform flights, but the sim-to-real claim leans on an unvalidated perfect-tracking assumption and the abstract oversells the lightweight board's speed.","tokens_in":14904,"tokens_out":1985,"would_cite":true,"duration_ms":23954,"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":"CORB-Planner claims that a single RL policy, trained in about ten minutes on a simplified obstacle course using only safe-flight-corridor observations, can plan real-time B-spline trajectories for quadrotors, hexarotors, LiDAR, and vision p","keywords":["reinforcement learning","safe flight corridor","B-spline trajectory planning","sim-to-real transfer","UAV autonomy","high-speed flight","onboard planning","SDCQ"],"falsifier":"Take the vmax=10 policy and fly it in the same cluttered environment while deliberately adding 100 ms of tracking latency or lowering the controller gains until the platform noticeably lags the B-spline. If success rate collapses while the planning inputs are unchanged, the perfect-tracking premise is the load-bearing link. A cleaner computational test: run a single realistic quadrotor simulator with imperfect tracking but the identical SFC observations and measure whether the policy's success rate falls below the no-physics training result.","tokens_in":14058,"feed_emoji":"🚁","tokens_out":6795,"duration_ms":79460,"temperature":0.7,"pith_summary":"CORB-Planner sets out to solve the sim-to-real problem for RL-based UAV trajectory planning by making the observation space platform-independent: instead of raw LiDAR or depth images, the policy sees a safe flight corridor built around an A* reference path, plus recent B-spline control points and a time index. The policy outputs one acceleration control point per step, which becomes the next B-spline control point, so the planner produces a smooth, dynamically feasible trajectory that the UAV's low-level controller tracks. Training is done in a deliberately simplified simulation without physics, using an easy-to-hard wall course, an exploitation-decoupled sampling scheme, and a soft decomposed-critic Q (SDCQ) value algorithm; the paper reports convergence in about ten minutes. If correct, a single policy can transfer across airframes and sensor suites and run on a tiny onboard computer, reaching 8.2 m/s in cluttered real-world flights. The paper's central claim is that low-dimensional SFC observations, not sensor-level details, are the right substrate for reusable RL planners in high-speed flight.","feed_headline":"RL planner trained in 10 minutes flies drones at 8.2 m/s","feed_subtitle":"Safe-corridor observations let one simulation policy transfer to quadrotors, hexarotors, and vision or LiDAR drones.","key_machinery":"The central object is the safe flight corridor as a low-dimensional observation. SFC is built by pruning an A* path into a minimal polyline, then attaching left and right rectangular sub-corridors whose widths record the distance to the nearest obstacle on each side within a chosen z-band. Each sub-corridor is described by eight numbers, so a local window of nine sub-corridors plus recent control points yields a 66-dimensional state. The companion mechanism is B-spline continuation: the RL action is a bounded acceleration control point, mapped through a cylindrical normalization to enforce uniform acceleration limits, then integrated to produce the next position control point; the third-orde","core_discovery":"The central claim: a safe flight corridor (SFC) can be a compact, platform-independent observation for an RL policy that plans B-spline trajectories. Around an A* reference polyline, SFC is a sequence of rectangular sub-corridors, each described by eight numbers (endpoints, left/right obstacle distances, z-range), so under 100 features encode the obstacle layout. The policy takes the nine nearest sub-corridors plus three recent control points and time (66 dimensions) and emits an acceleration control point in a cylindrical, uniformly constrained space, which becomes the next B-spline control point. Training uses the SDCQ value algorithm, a difficulty-progressive course, and multiple explorat","pith_inferences":["If the SFC representation is the true source of generality, then an optimization-based planner given the same 66-dimensional corridor observation should also transfer across platforms; testing that would separate the contribution of the representation from the contribution of RL.","The perfect-tracking training assumption implies a sharp test: degrade the low-level tracker by adding latency or lowering gains and measure how success rate drops; the 8.2 m/s figure should be read as specific to the tested low-level controller, not a platform-independent ceiling.","Because the action space enforces uniform horizontal acceleration, the policy is likely to generalize to vehicles with different mass and rotor layouts, but vehicles with asymmetric actuation limits would require a different action transform; extending the cylindrical mapping to those cases is a natural next step.","The moving-obstacle experiments were limited to slow pedestrians; extending CORB-Planner to faster dynamic obstacles would require the SFC to encode velocity information, which the current 66-dimensional state does not include."],"forward_implications":["Deploying CORB-Planner on a new vehicle does not require retraining: the same agents trained at maximum velocities of 4, 5, 7, 10, and 15 m/s ran on quadrotors, a hexarotor, LiDAR, and vision platforms in the reported experiments.","Because the policy outputs only B-spline control points, onboard planning cost is small enough to run at 25 Hz on a 275 g quadrotor with a small ARM board, leaving most CPU capacity to odometry.","Training is fast and cheap, about ten minutes on a desktop GPU without physics simulation, which could make RL planning practical for rapid iteration in field robotics.","On dense-wall courses, the aggressive variant of CORB-Planner is faster but less reliable than the conservative variant, so choosing reward weights trades speed against success rate.","Slowly moving obstacles can be handled by updating the occupancy grid and commanding a hover when no A* path exists, as shown in the moving-forest test."],"fun_headline_variants":["Corridor-based RL planner trains in 10 min, flies at 8.2 m/s","RL drone planner uses corridors to fly 8.2 m/s after 10-min training","10-minute RL training, corridor observations, 8.2 m/s drone flight","Safe-corridor observations let RL fly drones at 8.2 m/s with 10-min training"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing assumption is that any UAV with a capable low-level controller tracks the planned B-splines almost perfectly, because the training simulation omits physics entirely and says 'the tracking of B-spline trajectories is assumed to be perfectly precise' (Section IV-B); if tracking lag or platform-specific dynamics violate this, the ten-minute sim-to-real transfer and the 8.2 m/s results do not follow.","fun_headline_variants_meta":{"raw":{"variants":["Corridor-based RL planner trains in 10 min, flies at 8.2 m/s","RL drone planner uses corridors to fly 8.2 m/s after 10-min training","10-minute RL training, corridor observations, 8.2 m/s drone flight","Safe-corridor observations let RL fly drones at 8.2 m/s with 10-min training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00088,"raw_usage":{"total_tokens":3666,"prompt_tokens":793,"completion_tokens":2873,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2777}},"tokens_in":537,"tokens_out":2873,"duration_ms":22872,"temperature":1.0,"reasoning_tokens":2777,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T16:51:06.182813+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the vmax=10 policy and fly it in the same cluttered environment while deliberately adding 100 ms of tracking latency or lowering the controller gains until the platform noticeably lags the B-spline. If success rate collapses while the planning inputs are unchanged, the perfect-tracking premise is the load-bearing link. A cleaner computational test: run a single realistic quadrotor simulator with imperfect tracking but the identical SFC observations and measure whether the policy's success rate falls below the no-physics training result.","supporting_citations":[],"review_version":1}