{"id":"12c2cd57-17d4-480c-82b6-5ba084dcf2cf","arxiv_id":"2505.06923","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A single lightweight network replaces the cascade of detection, mapping, planning, and control, and tracks an uncooperative human target at up to 6 m/s in cluttered real-world environments.","lead":"This paper presents YOPOv2-Tracker, an end-to-end neural network that maps a quadrotor's camera images and flight state directly to thrust and attitude commands for chasing a moving target. The system was tested on a small drone in real forests and buildings, reaching tracking speeds of up to 6 m/s with only about 8 ms of perception-to-action latency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training loss (Eq. 16) contains no actuator-feasibility term, so the Hermite trajectories defined by Eqs.","rationale":"The reader's weakest assumption identifies the right soft spot. The paper is otherwise coherent: the differentiable-cost training in Eq. (16) is a real contribution, and the real-world flights are meaningful evidence that the system works at 6 m/s in the demonstrated runs. But the headline claim and the quantitative success-rate comparisons presuppose that every selected trajectory can be physically executed. The method deliberately removes the position controller and maps desired acceleration directly to attitude. The disturbance observer in Eq. (15) estimates d and shifts the desired acceleration, but it does not change the fact that collective thrust is bounded. When the Hermite acceleration exceeds the achievable acceleration, the observer will simply estimate the deficit as disturbance and feed it back; this cannot restore tracking, and may saturate or oscillate. This is not a disagreement with external consensus but an internal gap between what is optimized (smooth, safe, goal-reaching trajectories) and what the safety argument requires (actuator-feasible trajectories). The paper has no feasibility cost, and the tanh bounding in Eq. (7) constrains only endpoint derivatives, not the polynomial interior. The proposed actuator-saturation test directly probes whether this gap matters in the claimed operating regime. If it does, the natural remedy is a feasibility penalty or pre-execution validation of the selected primitive, consistent with a CONDITIONAL verdict. Since the reader already recommended CONDITIONAL on essentially this basis, I do not propose moving the verdict.","tokens_in":23138,"tokens_out":3733,"duration_ms":40969,"concrete_test":"Re-run the Section IV-B tracking benchmark at 7 m/s with the same 10 trials, but insert a hard actuator-saturation block between the desired thrust/attitude computation and the PX4 attitude controller, limiting total acceleration to 4.7g per the platform's stated thrust-to-weight ratio. Log the fraction of executed trajectories for which ||f_ddot(t)|| computed from Eq. (2) exceeds 4.7g at any discretized t in [0,T]. If the success rate or collision/emergency-stop statistics differ from the saturation-free Fig. 10B/C, the unmodeled feasibility assumption is load-bearing; if the results are identical, the concern is not practically limiting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on the network-generated fifth-order polynomial trajectories being executable by the physical quadrotor. The training objective L in Eq. (16) is a weighted sum of smoothness J_s, safety J_c, goal J_g, and detection losses; no term penalizes violation of actuator limits or dynamic feasibility. Endpoint derivatives are bounded by tanh in Eq. (7), and T is fixed to 2r/(alpha*v_max), but the acceleration profile of the Hermite polynomial between the actual initial state and the predicted endpoint is not bounded by these quantities. For a fifth-order polynomial with boundary derivatives in [-alpha*v_max, alpha*v_max] and [-alpha^2*a_max, alpha^2*a_max], the interior acceleration can substantially overshoot the endpoint values, especially when the current state has large velocity or acceleration and the refined radius r' is small. Section III-D's disturbance observer estimates the lumped disturbance d from state and last commands and adds it to the desired acceleration, but it cannot create thrust beyond the collective limit; if the planner commands more than the 4.7:1 thrust-to-weight ratio allows, the attitude loop saturates, and the claimed safety and tracking performance in Section IV degrade. The 8 m/s navigation experiment reports attitude accelerations mostly within 5 m/s^2, but that is an average operating point, not a certificate over the primitive library or over tracking at 6-7 m/s with an evasive target.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes YOPOv2-Tracker, an end-to-end quadrotor tracking and navigation framework that directly maps RGB-D images and onboard state estimates to thrust and attitude commands. The architecture is a single fully convolutional network that predicts offsets and costs for a set of motion primitives, along with objectness scores and target positions, thereby integrating detection, search, and trajectory optimization into one forward pass. Training is performed by back-propagating differentiable trajectory costs (smoothness, safety, goal) and detection losses through the network without expert demonstrations or reinforcement learning. At deployment, a disturbance observer compensates for model uncertainty, and an EKF-based target estimator drives the yaw. The authors report simulation comparisons with Elastic Tracker and Visibility-aware Tracker, plus real-world flights in dense forests and buildings at tracking speeds up to 6 m/s and navigation speeds up to 8 m/s.","tokens_in":23441,"tokens_out":3581,"duration_ms":36571,"significance":"If the claimed performance holds, this is a notable contribution: the system demonstrates that a single lightweight network can replace the traditional detection-mapping-planning-control cascade while preserving interpretability through primitive anchors and explicit trajectory costs. The real-world flights with an uncooperative human target in cluttered environments are credible evidence of practical utility, and the low reported latency (3 ms in simulation, 8.2 ms onboard) is an important engineering result. The code-release commitment strengthens reproducibility. However, the quantitative validation of smoothness and safety is partially circular because the reported metrics are the same cost functions used in the training objective, and the training objective lacks a dynamic-feasibility term, leaving a gap between the planned polynomial trajectories and the physical actuator limits. These issues reduce confidence in the quantitative comparisons until addressed, but they do not invalidate the central design contribution.","major_comments":[{"comment":"The smoothness and safety metrics reported in Fig. 10E-F are the very cost functions J_s and J_c that appear as training losses in Eq. (16), and the predicted cost y_c used for trajectory selection is supervised by the same quantities via Eq. (25). Therefore the finding that YOPOv2 achieves better smoothness and safety than Elastic Tracker and Visibility-aware Tracker is partly by construction: the network is explicitly trained to minimize these integrals, whereas the baselines optimize different objectives (e.g., hard visibility or corridor constraints). To substantiate the claimed superiority, please either report metrics that are not part of the training objective (for example, commanded-versus-actual acceleration, control effort, or the baseline methods' own cost definitions) or clearly acknowledge this circularity and soften the corresponding comparison.","section":"Sec. IV-B1, Fig. 10E-F; Eqs. (16), (17), (20), (25)"},{"comment":"The training loss in Eq. (16) contains no actuator-feasibility term. Equation (7) bounds only the endpoint derivatives through the tanh nonlinearity, and the execution time is fixed to T = 2r/(alpha v_max), but the interior acceleration of the fifth-order Hermite polynomial (Eq. (2)) can substantially exceed the endpoint values, especially when the current state has large velocity or acceleration and the refined radius r' is small. The disturbance observer of Eq. (15) can compensate for lumped disturbances but cannot increase collective thrust beyond the physical limit; if the desired wrench F from Eq. (13) exceeds the platform's 4.7:1 thrust-to-weight ratio, the attitude loop saturates and the safety and tracking claims in Section IV are not guaranteed. Figure 19 shows one 8 m/s navigation trial with attitude accelerations mostly within 5 m/s², but that is not a certificate over the primitive library or over 6-7 m/s tracking with an evasive target. Please provide an analysis of the maximum feasible acceleration over the primitive and offset space, or add a feasibility penalty or constraint to the training objective, and report commanded-versus-actual acceleration statistics across the real-world trials.","section":"Sec. III-C, III-D, III-E; Eqs. (7), (16); Sec. IV-C2"},{"comment":"The success-rate comparison in Fig. 10B is based on only 10 random forest realizations per target speed. With such a small sample and a stochastic evader whose goal is randomly switched, the reported differences at intermediate speeds may fall within sampling variability; no confidence intervals or statistical significance tests are provided. Since the central claim of superior success rates is load-bearing, please add more trials or report variance and confidence bounds to demonstrate that the differences are not due to chance.","section":"Sec. IV-B1, Fig. 10B"}],"minor_comments":[{"comment":"The abstract contains a typo: 'limitting' should be 'limiting'.","section":"Abstract"},{"comment":"The university name is misspelled as 'Egineering'; it should be 'Engineering'.","section":"Author affiliation"},{"comment":"The network input is described as '6-dimensional states (initial velocity and acceleration)', but the text later refers to 'state inputs concatenated with image features' without specifying whether the state includes other quantities such as the goal direction; please clarify the exact input vector.","section":"Sec. III-C1"},{"comment":"The phrase 'over 2 times faster than existing state-of-the-art methods in cluttered real-world scenarios' is based on comparing the maximum tracking speed (6 m/s) with the 2.5 m/s reported in [2]; please state the comparison explicitly and note that different platforms and environment difficulties may limit the directness of this comparison.","section":"Sec. IV-B3"},{"comment":"The notation for the binary indicators 1_pos and 1_neg is clear, but the handling of ignored samples in the loss could be stated more explicitly; the current text notes they are not considered for L_obj but does not clearly specify which trajectory losses apply to them.","section":"Sec. III-E, Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The central architecture and real-world demonstrations are credible and potentially significant. The main obstacles to acceptance are the circular evaluation metrics and the missing dynamic-feasibility analysis; both are addressable within the scope of a revision. If the authors can show that the generated trajectories respect actuator limits in the reported flight regimes and provide independent evaluation metrics, the paper could become a strong contribution to the field."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Things you should know about YOPOv2-Tracker before you spend time on it: the hardware results are the strongest part. A 155 mm quadrotor tracking a human at up to 6 m/s through a forest and 5.6 m/s through a building, fully onboard, at roughly 8 ms latency, is a real systems achievement. The method is also a genuine extension of the authors' prior YOPO work: it adds objectness prediction, target-position regression, EKF gating, and NMS to the one-stage planner. That is not a rebranding. What is new is the combination of detection-style primitive anchors with trajectory-cost backpropagation, and it is explained clearly enough to reproduce. The main soft spot is the evaluation, not the architecture. Figure 10E-F compares smoothness and safety costs against baselines using exactly the cost functions that define the training objective (Eqs. 17 and 20). That is expected by construction and should not be presented as independent evidence. The success-rate and latency numbers are independent of the training loss and are more convincing, but 10 trials per condition is thin. The claimed consistency at 7 m/s needs error bars or more trials. Code is promised but not yet available, so independent verification is not possible right now. The 'first end-to-end tracker' claim is plausible but never tested against D-VAT [43], which is cited but not compared; that omission should be fixed. The actuator-feasibility concern raised in the stress-test is legitimate but not fatal. The training loss has no term penalizing violation of thrust limits, and a Hermite polynomial can overshoot its tanh-bounded endpoint derivatives in the interior, especially with small refined radius and large initial velocity. That said, the real-world data in Fig. 19 show attitude accelerations mostly within 5 m/s squared at 8 m/s, which suggests the planner operates within the platform's 4.7:1 thrust-to-weight ratio in practice. Still, the authors should either add a feasibility penalty or discuss the worst case explicitly; as written, the safety claim in Section IV is not certified. Overall, this is a solid systems contribution. The training idea, backpropagating trajectory costs from a privileged ESDF instead of using expert demonstrations or RL, is interesting and well motivated. I would send it to a serious referee, with the instruction that the circular metrics, low trial count, and missing D-VAT comparison need to be addressed before acceptance. If your work touches learning-based planning or aerial tracking, it is worth a reading-group slot.","headline":"Real end-to-end tracking with credible flight demos, but the headline smoothness/safety metrics are the training loss itself and the feasibility gap is real.","tokens_in":751,"tokens_out":1023,"would_cite":false,"duration_ms":38068,"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 single fully convolutional network can turn RGB-D images and onboard state directly into attitude and thrust commands, enabling a small quadrotor to track an unpredictable human at up to 6 m/s through dense forests…","keywords":["end-to-end learning","aerial target tracking","quadrotor navigation","motion primitives","trajectory optimization","disturbance observer","vision-based control","RGB-D perception"],"falsifier":"During a 6 to 8 m/s forest or building run, record the commanded acceleration from the network's trajectory alongside the physical acceleration available from the thrust-to-weight ratio and the attitude controller's tracking error; if commanded accelerations regularly exceed the vehicle's capability, or if the disturbance observer's compensation diverges during sharp target turns, the core tracking claim fails.","tokens_in":22924,"feed_emoji":"🤖","tokens_out":6601,"duration_ms":64931,"temperature":0.7,"pith_summary":"The paper tries to show that 'less is more' works for agile aerial tracking. Instead of chaining detection, mapping, planning, and control, one lightweight fully convolutional network can map RGB-D images and flight state directly to attitude and thrust commands. If this holds, quadrotors with limited onboard sensing and computation could track fast, evasive, uncooperative targets in cluttered environments where traditional cascaded systems are too slow. The authors claim this is the first end-to-end tracker to achieve high-speed tracking of an unpredictable target in real-world clutter, demonstrated by chasing a human at up to 6 m/s through forests and buildings with a 155 mm quadrotor. The design keeps interpretability by making each image grid cell correspond to a motion primitive, so the network performs detection-like scoring and trajectory refinement rather than black-box action regression, and it is trained without expert demonstrations or reinforcement learning by back-propagating trajectory costs computed from a privileged map.","feed_headline":"One network turns pixels into drone thrust for 6 m/s tracking","feed_subtitle":"One learned forward pass replaces detection, mapping, planning, and control, cutting latency to 8.2 ms","key_machinery":"The load-bearing object is the primitive-anchored fully convolutional predictor. The image is divided into a 5 by 3 grid, and each cell corresponds to one motion primitive, a candidate endpoint on a sphere in the camera frustum described by polar angle, azimuth angle, and radius. The network predicts offsets and end derivatives that refine the primitive into a fifth-order Hermite polynomial with fixed execution time, plus a predicted cost, an objectness score, and the target position. During training, smoothness, safety, and goal costs are differentiated through the Hermite map to the network weights, so trajectory optimization and learning share one gradient path. At run time, the chosen trajectory's second derivative, corrected by a high-gain disturbance observer, is converted through differential flatness into desired attitude and thrust, removing the position controller and the reference-state mismatch that earlier end-to-end planners face.","core_discovery":"The central claim is that target tracking does not need the usual serial pipeline of detection, occupancy mapping, path search, trajectory optimization, and position control run one after another. A single fully convolutional network, called YOPOv2-Tracker, takes an RGB-D image and the current velocity and acceleration and outputs, for each of a grid of motion primitives spanning the camera frustum, refined offsets, endpoint derivatives, a trajectory cost, an objectness score, and a target position. The best primitive is turned into a fifth-order Hermite trajectory, and a disturbance observer converts the desired acceleration into thrust and attitude. Because mapping, planning, and detection are folded into one forward pass, the system responds in about 3 ms in simulation and 8.2 ms onboard, and in experiments it tracks an uncooperative human at up to 6 m/s through dense forest and buildings. The training signal is itself part of the claim: gradients of smoothness, safety, and goal costs from a privileged distance-to-obstacle map are back-propagated through the trajectory to the network weights, so no expert demonstrations or reinforcement-learning interactions are required.","pith_inferences":["The anchor-primitive mechanism could transfer to multi-target tracking or to ground robots, since the objectness head could be extended to multiple detections per grid cell.","The fixed execution time is a bottleneck; learning or adapting the execution time as a network output would likely improve agility during fast escapes and sharp target turns.","The recipe of using a differentiable cost map as the critic could be applied to other optimization-based robot tasks beyond flight, such as manipulation or racing, wherever a privileged cost can be computed.","The claim that target motion prediction is unnecessary could be tested directly by adding a short-horizon predictive head and comparing success rates; if prediction helps substantially, the 'unpredictable target' argument would need qualification."],"forward_implications":["Tracking success at target escape speeds of 3 to 7 m/s in simulation, and real-world human tracking at up to 6 m/s, becomes possible using only onboard RGB-D perception on a small quadrotor.","Perception-to-action latency drops to about 3 ms in simulation and 8.2 ms onboard, an order of magnitude faster than the compared cascaded trackers.","The same network, with the target-related outputs removed, acts as a high-speed navigation policy, so a single training recipe covers both tracking and obstacle avoidance.","Because the network outputs a chosen trajectory rather than a raw action, the system remains interpretable: each primitive corresponds to a feasible candidate within the camera's field of view.","Training by back-propagating privileged trajectory costs offers a third path between imitation learning and reinforcement learning for robot navigation tasks."],"supporting_citations":[{"why":"Serves as the hard-constraint baseline tracker whose spatio-temporal corridor formulation is compared against in latency, success rate, visibility, smoothness, and safety.","marker":"[1]"},{"why":"Serves as the visibility-aware soft-constraint baseline and provides the real-world maximum tracking speed that the paper claims to more than double.","marker":"[2]"},{"why":"Supplies the high-speed vision-based flight baseline and the argument that processing latency is critical, and is also compared against for navigation success rates.","marker":"[4]"},{"why":"Is the authors' prior one-stage planner that this work extends by adding detection, objectness scores, and tracking costs.","marker":"[6]"},{"why":"Provides the anchor-box detection design that motivates the use of motion primitives as trajectory anchors, and is used for real-world target annotation.","marker":"[10]"},{"why":"Supplies the gradient-based safety cost formulation using distance-to-obstacle fields that the training loss adapts.","marker":"[20]"},{"why":"Provides the polynomial trajectory optimization approach used as the evader's privileged planner and as context for the classical pipeline being replaced.","marker":"[22]"},{"why":"Represents the prior end-to-end tracking attempt that the paper contrasts as crude action regression limited to laboratory settings.","marker":"[43]"},{"why":"Supplies the motion-primitive trajectory library concept used to cover the feasible space with candidate endpoints.","marker":"[45]"},{"why":"Provides the smoothness cost as the integral of squared derivatives and the polynomial trajectory formulation used in the Hermite representation.","marker":"[46]"}],"fun_headline_variants":["One forward pass from pixels to thrust, 8 ms latency","Single network learns to fly through forest at 6 m/s","No expert data: trajectory costs train the tracker end-to-end","Less is more: single network replaces full tracking pipeline"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The trajectory the network generates is assumed to be physically flyable by the actual quadrotor, even though the training cost never checks whether the commanded acceleration stays within the motors' limits; if the planned path demands more acceleration than the attitude controller can provide, the safety and tracking guarantees degrade.","fun_headline_variants_meta":{"raw":{"variants":["One forward pass from pixels to thrust, 8 ms latency","Single network learns to fly through forest at 6 m/s","No expert data: trajectory costs train the tracker end-to-end","Less is more: single network replaces full tracking pipeline"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001433,"raw_usage":{"total_tokens":5829,"prompt_tokens":1045,"completion_tokens":4784,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":4715}},"tokens_in":661,"tokens_out":4784,"duration_ms":28868,"temperature":1.0,"reasoning_tokens":4715,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:30:11.975342+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"During a 6 to 8 m/s forest or building run, record the commanded acceleration from the network's trajectory alongside the physical acceleration available from the thrust-to-weight ratio and the attitude controller's tracking error; if commanded accelerations regularly exceed the vehicle's capability, or if the disturbance observer's compensation diverges during sharp target turns, the core tracking claim fails.","supporting_citations":[{"cited_title":"Elastic tracker: A spatio-temporal trajectory planner for flexible aerial tracking,","cited_arxiv_id":null,"evidence_quote":"Serves as the hard-constraint baseline tracker whose spatio-temporal corridor formulation is compared against in latency, success rate, visibility, smoothness, and safety."},{"cited_title":"Visibility-aware trajectory optimization with application to aerial tracking,","cited_arxiv_id":null,"evidence_quote":"Serves as the visibility-aware soft-constraint baseline and provides the real-world maximum tracking speed that the paper claims to more than double."},{"cited_title":"You only plan once: A learning-based one-stage planner with guidance learning,","cited_arxiv_id":null,"evidence_quote":"Is the authors' prior one-stage planner that this work extends by adding detection, objectness scores, and tracking costs."},{"cited_title":"Gradient-based online safe trajectory generation for quadrotor flight in complex environments,","cited_arxiv_id":null,"evidence_quote":"Supplies the gradient-based safety cost formulation using distance-to-obstacle fields that the training loss adapts."},{"cited_title":"D-vat: End- to-end visual active tracking for micro aerial vehicles,","cited_arxiv_id":null,"evidence_quote":"Represents the prior end-to-end tracking attempt that the paper contrasts as crude action regression limited to laboratory settings."},{"cited_title":"Motion primitives-based path planning for fast and agile exploration using aerial robots,","cited_arxiv_id":null,"evidence_quote":"Supplies the motion-primitive trajectory library concept used to cover the feasible space with candidate endpoints."},{"cited_title":"Polynomial trajectory planning for aggressive quadrotor flight in dense indoor environments,","cited_arxiv_id":null,"evidence_quote":"Provides the smoothness cost as the integral of squared derivatives and the polynomial trajectory formulation used in the Hermite representation."}],"review_version":1}