{"id":"d5b26e53-c829-4ec3-b45e-db0499ba8850","arxiv_id":"2505.00432","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A neural network controller trained in simulation runs directly on the PX4 flight controller's microcontroller and tracks a square path on a real quadrotor with behavior similar to simulation.","lead":"This paper shows a drone can be flown by a neural network running directly on its small flight controller chip, replacing the standard control software. The team trained the network in a simulator, converted it to run on the microcontroller, and tested it on a real quadrotor in a motion capture lab, achieving similar tracking in both.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single unquantified square-path flight plus acknowledged motor-model mismatch leaves the robust sim2real claim under-supported; stress-test with agile maneuvers is needed.","rationale":"The paper's durable contribution is the open-source PX4 module and deployment pipeline; that part does not depend on perfect simulation fidelity and is supported by the demonstrated real flight and measured inference latency (93.4 µs). The contested claim is 'robust sim2real transferability.' For that claim to hold, the training simulation must be a faithful enough model of the real platform, especially in the actuator dynamics that the policy directly commands. The authors acknowledge a known mismatch in motor time constants, and the evaluation consists of one square-path mission with no quantitative error metric or baseline comparison. This makes the transfer claim stronger than the data support. The proposed stress test—an agile trajectory with repeated trials—directly probes whether the motor-model mismatch matters. If the policy tracks an aggressive trajectory in simulation but degrades in real flight, then the sim2real claim is conditioned on the specific benign trajectory, and the verdict should remain CONDITIONAL with added requirements. If it tracks well, the concern is retired. This matches the reader's conditional acceptance; we see no reason to move the verdict.","tokens_in":5197,"tokens_out":4272,"duration_ms":46305,"concrete_test":"Fly the same LMF platform with the NN controller on an agile setpoint trajectory (e.g., a fast figure-eight or a sequence of large step inputs producing accelerations well above the square-path level) in both simulation and real flight, repeated at least 5 times. Compute position-tracking RMSE, max error, and the distribution of motor commands for both, and compare the real-vs-simulation gap to the gap observed on the square trajectory. If the gap grows substantially (e.g., the real tracking error increases much more than the simulated one, or motor-command variance diverges), then the motor-time-constant assumption is load-bearing and the broad transfer claim must be conditioned; if the gap stays small, the concern is mitigated. As a supporting check, measure increasing versus decreasing motor time constants on a test bench to confirm the assumed symmetry.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim, stated in Section IV-B, is that the neural controller 'is shown to achieve similar performances on the real platform as it did in simulation, indicating robust sim2real transferability.' The supporting evidence is Figure 3, a side-by-side plot from one simulated and one real square trajectory. There are no error bars, no repeated trials, and no comparison with the classical PX4 position controller, so 'similar' is not quantitatively established. In Section V, the authors write that 'we observe a larger variance between the motor commands across these two experiments' and attribute it to 'inaccurate estimates of the motor time constants used during training.' Section III-A.2 states that identical motor time constants were chosen for increasing and decreasing RPM setpoints. If the motor dynamics are indeed mis-modeled, the policy has been trained on a plant that differs from the real one in a component the authors themselves flag. The chosen square path at hover-like speeds may not excite the spool-up/spool-down asymmetry, so the observed position-tracking agreement does not demonstrate robust transfer to more agile maneuvers or other platforms. The load-bearing weakness is therefore not merely a missing baseline; it is that the evidence for the transfer claim is too thin precisely in the dimension (actuator dynamics) the authors identify as inaccurate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an open-source integration of neural-network-based controllers into the PX4 autopilot stack for embedded flight controllers. A policy trained in the Aerial Gym simulator is converted to TensorFlow Lite Micro and deployed as a custom PX4 module that replaces the classical position, velocity, attitude, rate, and mixer cascade, mapping position setpoints to normalized motor RPM commands. The authors report inference latency of 93.4 microseconds and demonstrate the controller on a square-path trajectory in both simulation and real flight, arguing that the similar tracking performance indicates robust sim2real transferability. The complete pipeline, including training code and the PX4 module, is open-sourced.","tokens_in":5403,"tokens_out":2330,"duration_ms":26713,"significance":"If the central claims are substantiated, the work provides a genuinely useful engineering contribution: a reproducible, flight-ready pipeline for putting learned controllers on low-cost, off-the-shelf flight controllers, with a clear safety path (switching back to classical mode) and measured computational costs. The open-sourcing of both the simulator integration and the PX4 module is a concrete strength that lowers the barrier for learning-based aerial control research. The paper's significance is, however, currently limited by an evaluation that is too thin to support the strong sim2real-transfer claim: a single simulated and a single real square-trajectory run, without quantitative tracking-error metrics, repeated trials, or comparison against the classical PX4 controller. The contribution is real but the evidence needs strengthening.","major_comments":[{"comment":"The central claim that the neural controller 'is shown to achieve similar performances on the real platform as it did in simulation, indicating robust sim2real transferability' is supported only by one side-by-side plot of a single square trajectory in simulation and in real flight. There are no quantitative tracking-error metrics (e.g., RMSE in position or velocity), no repeated trials, no error bars, and no statistical comparison. 'Similar' is therefore not established by the presented evidence. Please provide numerical tracking errors for both the simulated and real runs, report multiple trials (at least for the real flight), and state explicitly how the plots were obtained and whether they represent typical behavior.","section":"IV-B, Figure 3"},{"comment":"The authors themselves attribute the larger variance in motor commands between simulation and reality to inaccurate estimates of the motor time constants, and Section III-A.2 states that identical time constants were chosen for increasing and decreasing RPM setpoints. The demonstrated square path at relatively mild speeds may not excite the spool-up/spool-down asymmetry that the mis-modeled time constants would affect. Consequently, the observed position-tracking agreement does not by itself establish that the policy would transfer robustly to agile maneuvers or to platforms with different motor dynamics. Please add at least one more demanding trajectory (e.g., with rapid accelerations and aggressive setpoint changes) and, ideally, report motor time-constant estimates from a test bench to close the acknowledged modeling gap.","section":"III-A.2 and V"},{"comment":"The paper does not compare the neural controller's tracking performance against the classical PX4 position controller on the same flight platform and trajectory. Without such a baseline, the reader cannot judge whether the observed absolute tracking quality is acceptable, nor whether the sim2real agreement is distinctive of the learned policy. Adding a baseline comparison (classical controller under the same motion-capture and setpoint conditions) would make the performance claim much more interpretable and would directly strengthen the paper's central argument.","section":"IV-B"}],"minor_comments":[{"comment":"The reward function is only described qualitatively as containing terms for position, orientation, velocities, action magnitudes, and action differences. Please provide the actual reward weights or a reference to the specific Aerial Gym configuration used, since the trained policy's behavior cannot be reproduced without them.","section":"III-A.3"},{"comment":"The module's scheduling on the angular velocity topic at 650 Hz is mentioned, but it is unclear what the actual control-loop frequency is after pre- and post-processing (137.6 microseconds) given the 650 Hz trigger. Please state the loop rate achieved and clarify the relationship between inference time, loop time, and the topic rate.","section":"III-C.1"},{"comment":"Figure 3 would benefit from labeled axes, units, and a legend distinguishing simulated versus real trajectories, as well as clear markers for the start and end points of the square. In the current manuscript text the figure is referenced only as a side-by-side plot, and the visual comparison is hard to quantify.","section":"IV-B"},{"comment":"The introduction contains a typo, 'auopilot' instead of 'autopilot', and the phrase 'to provide an off-the-shelf solution' would read better with a comma after 'solution'. Minor editorial issues throughout the text should be cleaned up.","section":"II"},{"comment":"The description of the platform states that the NVIDIA Orin NX is used only for relaying motion-capture pose. It would be useful to state whether the pose is sent to the flight controller over a wired or wireless link and at what rate, since the controller depends on this external state source.","section":"IV-A"}],"recommendation":"major_revision","confidential_remarks":"The paper is more of a systems and open-source contribution than a new algorithmic result, which is fine if the journal values reproducible deployment pipelines. The main question for the editor is whether the current evaluation depth (a single square trajectory with no quantitative metrics) is acceptable for the journal's standards; in my view it is not yet, but the issues are addressable with additional experiments and analysis rather than being fundamental."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline: this is a real, open-source engineering contribution—a PX4 module that runs a small neural network directly on the flight controller's STM32, replacing the standard position-through-actuation cascade, with a real flight demonstration. That is new relative to the prior work it cites, which mostly relies on a laptop or an Orin for inference. The code and conversion pipeline are shipped, and the module integrates with uORB and keeps the classical PX4 modes available for safe takeoff and landing. For anyone who wants to deploy learned controllers on multirotors without companion compute, this is a genuinely useful enabler.\n\nThe paper's central claim is that the policy transfers from Aerial Gym to the real platform robustly. That claim is under-supported. The evidence is one square trajectory in simulation and one in reality, plotted side by side, with no error bars, no repeated trials, no quantitative tracking error, and no comparison against PX4's own classical controller. \"Similar\" is doing a lot of work. The authors also admit in the Discussion that the motor commands differ more between sim and real, and they attribute that to inaccurate motor time constants—the same component they chose symmetrically in simulation (Section III-A.2). That is not a fatal flaw, but it means the observed agreement on a gentle square path doesn't yet establish robust transfer to agile maneuvers, which is precisely where a spool-up/spool-down mismatch would show.\n\nI'd also be careful with the phrase \"flight-ready\" in the abstract. With one untripled, unquantified flight, the pipeline is ready for the community to test, but the controller's readiness claim is stronger than the data.\n\nWhat's good: the design choices are sound. Using TFLM on the MCU, keeping classical modes for safety, and reporting inference latency (93 µs) all suggest the authors thought about real deployment constraints. The limitation section is honest, not buried. The open-source repos give the community something concrete to build on. The work is coherent on its own terms; the gap is evaluation, not logic.\n\nWho is this for: the aerial learning/control community and anyone working on TinyML on flight controllers. It deserves a serious referee. My recommendation is to send it to review, and to require a tightened evaluation—repeated flights, error statistics, a classical baseline, and ideally a more aggressive maneuver or an identified motor model—before it can be accepted with the robust-transfer claim intact.","headline":"A useful, genuinely open-sourced PX4 module for onboard neural control, with a real flight demo, but the sim2real evidence is a single unquantified square path and the robust-transfer claim is overstated.","tokens_in":5964,"tokens_out":2738,"would_cite":true,"duration_ms":28550,"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":"A neural-network position controller can replace PX4's control cascade and run on the flight controller's own chip, matching simulation in real flight.","keywords":["neural network control","PX4 autopilot","embedded flight controller","TensorFlow Lite Micro","sim-to-real transfer","reinforcement learning","quadrotor","Aerial Gym Simulator"],"falsifier":"Put the motor on a test bench, measure how quickly RPM rises and falls, and compare with the single time constant used in training; also fly an aggressive maneuver in simulation and in reality to see if the near-match in position tracking breaks down.","tokens_in":4973,"feed_emoji":"🚁","tokens_out":7110,"duration_ms":65601,"temperature":0.7,"pith_summary":"The paper sets out to show that a neural-network controller trained in simulation can be deployed as a normal flight mode inside the PX4 autopilot, running directly on the flight controller's microcontroller rather than on a companion computer. A small fully connected policy replaces the classical cascade of position, velocity, attitude, rate, and mixing controllers, producing normalized motor RPM setpoints directly from state observations. Flight tests on a 1.2 kg quadcopter show the policy tracks a square trajectory about as well in the real world as it does in simulation, while the motor commands show more scatter that the authors trace to estimated motor time constants. The value of the claim is practical: if it holds, learning-based control policies can be tested on off-the-shelf autopilots without custom low-level firmware or extra compute hardware.","feed_headline":"Neural controller flies on embedded PX4 hardware, matching simulation","feed_subtitle":"A PPO-trained policy runs at 650 Hz on the flight controller chip, opening embedded hardware to learned control.","key_machinery":"The load-bearing mechanism is the neural control module itself: a PX4 flight-mode module that reads the same uORB state topics as the classical controllers and publishes actuator commands in their place, so the neural network runs as a drop-in replacement for the position, velocity, attitude, body-rate, and mixer stages. The policy is converted from PyTorch to a TensorFlow Lite Micro C-array by a script in the simulator repository, and the module is triggered by the angular-velocity topic, which updates at 650 Hz on the Pixracer Pro. This topic-swapping design is what lets the same module be adapted later to replace only selected parts of the cascade, such as control allocation alone.","core_discovery":"The central claim is that an end-to-end neural position controller can substitute for the whole classical control cascade inside PX4 and still fly a real quadcopter with accuracy comparable to simulation. The policy is a two-hidden-layer network (64 and 32 ReLU neurons) with 15 state inputs and four outputs, trained with PPO in the Aerial Gym Simulator, converted to TensorFlow Lite Micro, and compiled into the autopilot firmware as a dedicated flight mode. On the STM32H743-based Pixracer Pro, inference takes about 93.4 µs and the full control loop 137.6 µs, with the module scheduled at the 650 Hz angular-velocity topic. The real vehicle followed the same square setpoint path as the simulated one, supporting the paper's claim of reliable sim-to-real transfer for low-level motor control.","pith_inferences":["Beyond the paper, the same topic-swapping module could serve as a template for neural observers or for switching among multiple specialized policies mid-flight, since the autopilot's safety checks remain active.","Beyond the paper, if the sim-to-real match holds under aggressive maneuvers, embedded neural control could address regimes where classical cascades are tuned conservatively, such as rapid attitude changes or per-rotor saturation.","Beyond the paper, a direct test of the authors' stated cause would be to retrain with measured asymmetric motor time constants and check whether the motor-command variance between sim and real shrinks."],"forward_implications":["Researchers can deploy neural control policies on off-the-shelf PX4 flight controllers without a companion computer.","Because neural control is a separate flight mode, operators retain classical PX4 controllers for takeoff, landing, and emergency switching.","The short inference time leaves capacity for higher-rate control or larger networks on the same microcontroller.","Measuring the motor time constants on a test bench should reduce the observed sim-to-real difference in motor commands.","The topic-based design means selected parts of the control cascade, such as the mixer, could be replaced by neural networks independently."],"supporting_citations":[{"why":"PX4 autopilot provides the open-source flight stack, safety checks, and hardware support that the neural module extends.","marker":"[1]"},{"why":"TensorFlow Lite Micro supplies the inference engine that lets the network run on the flight controller's microcontroller.","marker":"[2]"},{"why":"Aerial Gym Simulator is the training environment used to learn the position-tracking policy.","marker":"[3]"},{"why":"Describes the Learning-based Micro Flyer platform used for the real-world experiments.","marker":"[6]"},{"why":"Isaac Gym is the underlying physics simulation engine that Aerial Gym uses for highly parallelized policy training.","marker":"[9]"},{"why":"RL Games provides the open-source PPO implementation used to train the policy.","marker":"[10]"}],"fun_headline_variants":["Neural network takes over PX4 flight control, matches sim","Embedded NN flies real drone, sim-to-real at 650 Hz","Open-source neural flight mode for PX4 runs on-chip","Neural policy flies real quadcopter at 650 Hz on MCU","PX4 gets a neural autopilot that flies like simulation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulator's motor model is accurate enough, in particular that motor RPM rises and falls at the same rate; if that is wrong, the good match between simulated and real position tracking may not survive more aggressive flight.","fun_headline_variants_meta":{"raw":{"variants":["Neural network takes over PX4 flight control, matches sim","Embedded NN flies real drone, sim-to-real at 650 Hz","Open-source neural flight mode for PX4 runs on-chip","Neural policy flies real quadcopter at 650 Hz on MCU","PX4 gets a neural autopilot that flies like simulation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000389,"raw_usage":{"total_tokens":2045,"prompt_tokens":937,"completion_tokens":1108,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":553,"completion_tokens_details":{"reasoning_tokens":1016}},"tokens_in":553,"tokens_out":1108,"duration_ms":7856,"temperature":1.0,"reasoning_tokens":1016,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:41:20.260357+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Put the motor on a test bench, measure how quickly RPM rises and falls, and compare with the single time constant used in training; also fly an aggressive maneuver in simulation and in reality to see if the near-match in position tracking breaks down.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"PX4 autopilot provides the open-source flight stack, safety checks, and hardware support that the neural module extends."},{"cited_title":"Meier, D","cited_arxiv_id":null,"evidence_quote":"TensorFlow Lite Micro supplies the inference engine that lets the network run on the flight controller's microcontroller."},{"cited_title":"Multi-Task Reinforcement Learning for Quadrotors","cited_arxiv_id":"2412.12442","evidence_quote":"Describes the Learning-based Micro Flyer platform used for the real-world experiments."},{"cited_title":"Neural Lander: Stable Drone Landing Control using Learned Dynamics","cited_arxiv_id":"1811.08027","evidence_quote":"Isaac Gym is the underlying physics simulation engine that Aerial Gym uses for highly parallelized policy training."}],"review_version":1}