A Neural Network Mode for PX4 on Embedded Flight Controllers

Kostas Alexis; Mihir Kulkarni; Sindre M. Hegre; Welf Rehberg

REVIEW 3 major objections 5 minor 12 references

A Neural Network Mode for PX4 on Embedded Flight Controllers

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2505.00432 v1 pith:TQJBSQDX submitted 2025-05-01 cs.RO

Sindre M. Hegre , Welf Rehberg , Mihir Kulkarni , Kostas Alexis This is my paper · ORCID

classification cs.RO

keywords neuralnetworkcontrolPX4autopilotembeddedflightcontrollerTensorFlowLiteMicrosim-to-realtransferreinforcementlearningquadrotorAerialGymSimulator

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

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.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the 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.

The 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.

I'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.

What'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.

Who 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.

Referee Report

3 major / 5 minor

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.

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 (3)

[IV-B, Figure 3] 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.
[III-A.2 and V] 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.
[IV-B] 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.

minor comments (5)

[III-A.3] 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.
[III-C.1] 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.
[IV-B] 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.
[II] 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.
[IV-A] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the sim2real comparison is an empirical validation; self-citations are infrastructure, not load-bearing derivation.

full rationale

The paper makes no claimed derivation that reduces to its inputs. The central claim is that a neural-network controller trained in the Aerial Gym Simulator, deployed through a custom PX4 module, achieves similar tracking performance in simulation and on the real platform (Section IV-B, Figure 3). This is a direct empirical comparison between two independently measured executions, not a quantity obtained by fitting a model to the same outcome being predicted. The system-identification step (Section III-A.2) calibrates mass, inertia, thrust coefficient, and motor time constants of the simulator; the real-world flight result is not used to adjust these parameters, and the paper explicitly acknowledges the residual mismatch in motor commands and attributes it to inaccurate motor time constants (Section V). That is an honest limitation about model fidelity, not circularity. The self-citations to Aerial Gym Simulator [3] and the LMF platform [6] are used as the training environment and test platform; they do not supply a uniqueness theorem, an ansatz, or a forced conclusion. No equation in the paper defines a predicted quantity in terms of the same quantity from data. Concerns about thin evidence (single square-path trial, no error bars, no classical-PX4 baseline) bear on the strength of the sim2real claim, not on whether the claim is circular. The derivation chain is therefore self-contained as a pipeline demonstration, and no circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central sim2real claim rests on the system identification and simulation fidelity, as well as the external motion capture setup in the real flight. No new physical entities are introduced.

free parameters (5)

Thrust coefficient = not reported
Calculated by measuring motor RPM during hover and relating to steady-state force (Section III-A.2).
Thrust-to-torque ratios = not reported
Approximated in Section III-A.2.
Motor time constants = not reported
Identical constants chosen for increasing and decreasing RPM set points (Section III-A.2); identified as inaccurate in the Discussion.
Inertia estimates = from CAD, not data-fitted
Estimated using a CAD model (Section III-A.2).
Reward weights = tuned
Reward function tuned to enhance policy tracking in simulation (Section III-A.3).

assumptions (4)

domain assumption The Isaac Gym physics simulation faithfully represents the real quadrotor dynamics after system identification.
The policy is trained entirely in this simulator and then deployed on the real platform (Section III-A).
domain assumption PPO training in RL Games produces a policy that generalizes to the real world.
The authors rely on the trained policy transferring across the sim-to-real gap (Section III-A.3).
standard math TensorFlow Lite Micro conversion preserves the trained network's numerical behavior on the MCU.
The pipeline assumes that the converted C-array inference matches the PyTorch network outputs (Section III-C).
domain assumption Motion capture pose estimates provide ground-truth state for the policy inputs.
Real-world experiments use an external Qualisys motion capture system relayed through an Orin NX (Sections IV-A, IV-B).

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Neural Network Mode for PX4 on Embedded Flight Controllers." pith.science (2026). https://pith.science/paper/TQJBSQDX

@misc{pith2026250500432,
  author       = {Pith},
  title        = {Pith review of: A Neural Network Mode for PX4 on Embedded Flight Controllers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TQJBSQDX}},
  note         = {Machine review of arXiv:2505.00432}
}

read the original abstract

This paper contributes an open-sourced implementation of a neural-network based controller framework within the PX4 stack. We develop a custom module for inference on the microcontroller while retaining all of the functionality of the PX4 autopilot. Policies trained in the Aerial Gym Simulator are converted to the TensorFlow Lite format and then built together with PX4 and flashed to the flight controller. The policies substitute the control-cascade within PX4 to offer an end-to-end position-setpoint tracking controller directly providing normalized motor RPM setpoints. Experiments conducted in simulation and the real-world show similar tracking performance. We thus provide a flight-ready pipeline for testing neural control policies in the real world. The pipeline simplifies the deployment of neural networks on embedded flight controller hardware thereby accelerating research on learning-based control. Both the Aerial Gym Simulator and the PX4 module are open-sourced at https://github.com/ntnu-arl/aerial_gym_simulator and https://github.com/SindreMHegre/PX4-Autopilot-public/tree/for_paper. Video: https://youtu.be/lY1OKz_UOqM?si=VtzL243BAY3lblTJ.

Figures

Figures reproduced from arXiv: 2505.00432 by the authors.

**Figure 2.** Left: The Learning-based Micro Flyer platform. Right: The PX4 control cascade. The NN in the Neural Control [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Left: Performance of the neural network in simulation. Right: Performance of the neural network on the real platform. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 6 canonical work pages

[1]

11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorblockAtopspace 0.0ex \@IEEEauthorblockNinterlinespace 2.6ex \@IEEEauthorblockAinte...

work page
[2]

Meier, D

L. Meier, D. Honegger, and M. Pollefeys, `` en PX4 : A node-based multithreaded open source robotics framework for deeply embedded platforms ,'' in en 2015 IEEE International Conference on Robotics and Automation ( ICRA ) . 1em plus 0.5em minus 0.4em Seattle, WA, USA: IEEE, May 2015, pp. 6235--6240. [Online]. Available: http://ieeexplore.ieee.org/document...

work page arXiv 2015
[3]

David, J

R. David, J. Duke, A. Jain, V. J. Reddi, N. Jeffries, J. Li, N. Kreeger, I. Nappier, M. Natraj, S. Regev, R. Rhodes, T. Wang, and P. Warden, `` TensorFlow Lite Micro : Embedded Machine Learning on TinyML Systems ,'' Mar. 2021, arXiv:2010.08678. [Online]. Available: http://arxiv.org/abs/2010.08678

arXiv 2021
[4]

Kulkarni, W

M. Kulkarni, W. Rehberg, and K. Alexis, ``Aerial Gym Simulator : A Framework for Highly Parallelized Simulation of Aerial Robots ,'' Mar. 2025, arXiv:2503.01471 [cs]. [Online]. Available: http://arxiv.org/abs/2503.01471

arXiv 2025
[5]

Eschmann, D

J. Eschmann, D. Albani, and G. Loianno, ``Learning to Fly in Seconds ,'' Apr. 2024, arXiv:2311.13081. [Online]. Available: http://arxiv.org/abs/2311.13081

arXiv 2024
[6]

J. Xing, I. Geles, Y. Song, E. Aljalbout, and D. Scaramuzza, ``Multi- Task Reinforcement Learning for Quadrotors ,'' Dec. 2024, arXiv:2412.12442 [cs]. [Online]. Available: http://arxiv.org/abs/2412.12442

work page Pith review arXiv 2024
[7]

Kulkarni and K

M. Kulkarni and K. Alexis, ``Reinforcement Learning for Collision -free Flight Exploiting Deep Collision Encoding ,'' Feb. 2024, arXiv:2402.03947. [Online]. Available: http://arxiv.org/abs/2402.03947

arXiv 2024
[8]

Bauersfeld, E

L. Bauersfeld, E. Kaufmann, P. Foehn, S. Sun, and D. Scaramuzza, `` NeuroBEM : Hybrid Aerodynamic Quadrotor Model ,'' in Robotics: Science and Systems XVII , Jul. 2021. [Online]. Available: http://arxiv.org/abs/2106.08015

arXiv 2021

Show all 12 references

[9]

G. Shi, X. Shi, M. O'Connell, R. Yu, K. Azizzadenesheli, A. Anandkumar, Y. Yue, and S.-J. Chung, ``Neural Lander : Stable Drone Landing Control using Learned Dynamics ,'' in 2019 International Conference on Robotics and Automation ( ICRA ) , May 2019, pp. 9784--9790, arXiv:181...

2019 arXiv
[10]

Makoviychuk, L

V. Makoviychuk, L. Wawrzyniak, Y. Guo, M. Lu, K. Storey, M. Macklin, D. Hoeller, N. Rudin, A. Allshire, A. Handa, and G. State, ``Isaac Gym : High Performance GPU - Based Physics Simulation For Robot Learning ,'' Aug. 2021, arXiv:2108.10470 [cs]. [Online]. Available: http://ar...

2021 arXiv
[11]

Makoviichuk, ``Denys88/rl\_games,'' Mar

D. Makoviichuk, ``Denys88/rl\_games,'' Mar. 2025, original-date: 2019-01-13T05:35:44Z. [Online]. Available: https://github.com/Denys88/rl_games

2025
[12]

[Online]

`` en-US Open Source Autopilot for Drones ,'' uRL: https://px4.io/. [Online]. Available: https://px4.io/

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.

[1] [1]

11em plus .33em minus .07em @technote 4000 4000 100 4000 4000 500 `\.=1000 = #1 #1 #1 0pt [0pt][0pt] #1 * \| ** #1 \@IEEEauthorblockNstyle \@IEEEauthorblockAstyle \@IEEEauthordefaulttextstyle \@IEEEauthorblockconfadjspace -0.25em \@IEEEauthorblockNtopspace 0.0ex \@IEEEauthorblockAtopspace 0.0ex \@IEEEauthorblockNinterlinespace 2.6ex \@IEEEauthorblockAinte...

work page

[2] [2]

Meier, D

L. Meier, D. Honegger, and M. Pollefeys, `` en PX4 : A node-based multithreaded open source robotics framework for deeply embedded platforms ,'' in en 2015 IEEE International Conference on Robotics and Automation ( ICRA ) . 1em plus 0.5em minus 0.4em Seattle, WA, USA: IEEE, May 2015, pp. 6235--6240. [Online]. Available: http://ieeexplore.ieee.org/document...

work page arXiv 2015

[3] [3]

David, J

R. David, J. Duke, A. Jain, V. J. Reddi, N. Jeffries, J. Li, N. Kreeger, I. Nappier, M. Natraj, S. Regev, R. Rhodes, T. Wang, and P. Warden, `` TensorFlow Lite Micro : Embedded Machine Learning on TinyML Systems ,'' Mar. 2021, arXiv:2010.08678. [Online]. Available: http://arxiv.org/abs/2010.08678

arXiv 2021

[4] [4]

Kulkarni, W

M. Kulkarni, W. Rehberg, and K. Alexis, ``Aerial Gym Simulator : A Framework for Highly Parallelized Simulation of Aerial Robots ,'' Mar. 2025, arXiv:2503.01471 [cs]. [Online]. Available: http://arxiv.org/abs/2503.01471

arXiv 2025

[5] [5]

Eschmann, D

J. Eschmann, D. Albani, and G. Loianno, ``Learning to Fly in Seconds ,'' Apr. 2024, arXiv:2311.13081. [Online]. Available: http://arxiv.org/abs/2311.13081

arXiv 2024

[6] [6]

J. Xing, I. Geles, Y. Song, E. Aljalbout, and D. Scaramuzza, ``Multi- Task Reinforcement Learning for Quadrotors ,'' Dec. 2024, arXiv:2412.12442 [cs]. [Online]. Available: http://arxiv.org/abs/2412.12442

work page Pith review arXiv 2024

[7] [7]

Kulkarni and K

M. Kulkarni and K. Alexis, ``Reinforcement Learning for Collision -free Flight Exploiting Deep Collision Encoding ,'' Feb. 2024, arXiv:2402.03947. [Online]. Available: http://arxiv.org/abs/2402.03947

arXiv 2024

[8] [8]

Bauersfeld, E

L. Bauersfeld, E. Kaufmann, P. Foehn, S. Sun, and D. Scaramuzza, `` NeuroBEM : Hybrid Aerodynamic Quadrotor Model ,'' in Robotics: Science and Systems XVII , Jul. 2021. [Online]. Available: http://arxiv.org/abs/2106.08015

arXiv 2021