pith. sign in

arxiv: 2604.10262 · v1 · submitted 2026-04-11 · ⚛️ physics.optics

Deep Photonic Reservoir Computer Meets UAV Control: An ultra-fast learning-based compensator for agile flight in confined space

Qinxiao Ma , Ruiqian Li , Cheng Wang , Yang Wang This is my paper

Pith reviewed 2026-05-10 15:38 UTC · model grok-4.3

classification ⚛️ physics.optics
keywords photonic reservoir computerUAV controlconfined space flightresidual force predictionlaser dynamicsnanosecond inferencePID compensationCFD simulation
0
0 comments X

The pith

A deep photonic reservoir computer predicts UAV residual forces with nanosecond inference after millisecond training.

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

Unmanned aerial vehicles lose performance in confined spaces from unmodeled forces such as ground and ceiling effects that standard controllers cannot handle. Conventional learning compensators require extensive historical data, suffer from slow training, and incur high computational cost during operation. This paper integrates a hardware deep photonic reservoir computer built from semiconductor laser dynamics and optical feedback to supply intrinsic temporal memory. Only a linear readout layer is trained via ridge regression, delivering residual-force predictions that match or exceed those of TCN and MLP baselines. These predictions are injected into a PID controller through a feedforward path, producing improved closed-loop tracking stability.

Core claim

The hardware-implemented deep photonic reservoir computer, driven by semiconductor laser dynamics and optical feedback, supplies intrinsic temporal memory without explicit historical inputs. Training reduces to milliseconds by fitting only a linear readout layer through ridge regression, while inference reaches nanosecond latency. High-fidelity CFD simulations of proximity-induced flows confirm that residual-force prediction accuracy is comparable to or better than TCN and MLP baselines. Feeding the predictions into a nonlinear feedback PID controller via a feedforward channel measurably improves closed-loop tracking stability for UAVs operating in confined spaces.

What carries the argument

Deep photonic reservoir computer architecture that harnesses semiconductor laser dynamics and optical feedback to embed temporal memory directly in the hardware.

If this is right

  • Training time for the compensator drops from hours to milliseconds by optimizing only the linear readout.
  • Inference latency falls to nanoseconds, enabling real-time feedforward compensation inside the control loop.
  • Closed-loop tracking stability improves for agile maneuvers near surfaces.
  • The same architecture extends to fluid environments more complex than the tested cases without retraining the reservoir itself.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Photonic hardware could support edge-based compensation on small drones where digital neural networks exceed power or latency budgets.
  • The intrinsic memory property may simplify controller design for other robotic platforms that encounter time-varying fluid forces.
  • Direct comparison of the photonic hardware against the CFD model on a physical test rig would isolate any gap between simulation and reality.

Load-bearing premise

CFD simulations faithfully reproduce all relevant proximity-induced aerodynamic flows, and the photonic hardware can be built without noise, drift, or fabrication variability that would spoil the claimed training and inference speeds.

What would settle it

Running the physical photonic reservoir computer hardware aboard a real UAV in a confined test environment and checking whether its force predictions and resulting flight stability match the CFD-simulated performance within the reported accuracy margins.

Figures

Figures reproduced from arXiv: 2604.10262 by Cheng Wang, Qinxiao Ma, Ruiqian Li, Yang Wang.

Figure 1
Figure 1. Figure 1: Framework of the deep PRC compensation system. The system includes three parts: training and ground-truth generation; deep PRC inference; [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic architecture of the deep PRC. Virtual neurons for the [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of residual force prediction in two flight cases: [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Closed-loop trajectory tracking in two cases: [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Radar-plot comparison of deep PRC, MLP, and TCN across training [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Prediction error of deep PRC, MLP, and TCN during flight. At [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Unmanned aerial vehicles (UAVs) operating in confined, cluttered environments face significant performance degradation due to nonlinear, time-varying unmodeled dynamics-such as ground/ceiling effects and wake recirculation-that are unaccounted for in traditional controllers. While learning based compensators (e.g., MLPs, TCNs, LSTMs) struggle with historical data dependency, vanishing gradients, and prohibitive computational costs, this work pioneers the integration of a deep photonic reservoir computer (PRC) with feedforward control to overcome these limitations. Harnessing semiconductor laser dynamics and optical feedback, our hardware implemented deep PRC architecture achieves intrinsic temporal memory without explicit historical inputs, while reducing training time from hours to milliseconds and slashing inference latency to nanoseconds. Reliable high-performance CFD simulations capturing proximity-induced flows demonstrate that deep PRC delivers residual-force prediction accuracy comparable to or exceeding TCN/MLP baselines, while training only a linear readout layer via ridge regression. By injecting these predictions into a nonlinear feedback PID controller via a feedforward channel, the framework significantly enhances closed-loop tracking stability in confined spaces. Essentially, this work establishes the first deep PRC-based lightweight, ultrafast solution for real-time UAV dynamic compensation, with promising extensibility to unseen scenarios with more complex fluid environments.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes integrating a deep photonic reservoir computer (PRC), implemented via semiconductor laser dynamics with optical feedback, as a feedforward compensator for residual forces on UAVs in confined spaces. CFD simulations generate proximity-induced flow data to train only the linear readout layer via ridge regression, claiming intrinsic temporal memory without explicit history inputs, training times reduced to milliseconds, inference latency to nanoseconds, and residual-force prediction accuracy comparable to or exceeding TCN/MLP baselines. These predictions are injected into a nonlinear PID controller to improve closed-loop tracking stability.

Significance. If the simulation-to-hardware mapping holds, the work offers a promising path to ultra-low-latency, low-power learning-based control for agile UAV flight by exploiting photonic reservoir dynamics. The approach avoids vanishing-gradient issues and heavy compute of recurrent networks while providing a concrete engineering demonstration of PRCs in a closed-loop setting. The CFD-driven validation framework is a positive element that grounds the claims in physically motivated data.

major comments (2)
  1. [Abstract] Abstract: the central claims of 'comparable or exceeding' accuracy to TCN/MLP baselines and 'dramatic speed gains' (training from hours to milliseconds, inference to nanoseconds) are stated without any reported quantitative metrics, RMSE values, error bars, training/validation curves, or details on CFD data partitioning and cross-validation, leaving the performance assertions unsupported by evidence in the manuscript.
  2. [PRC hardware model section] PRC hardware model section: no analysis or Monte-Carlo simulation is provided to quantify how laser intensity noise, phase drift, or fabrication tolerances in the optical feedback loop alter the reservoir state trajectory, eigenvalue spectrum, or prediction accuracy relative to the ideal numerical model; because closed-loop UAV stability results are obtained by feeding the simulated PRC outputs directly into the PID, any such degradation would directly undermine the claimed stability improvement and the assertion that millisecond training and nanosecond inference are preserved in flight.
minor comments (1)
  1. [Abstract] The abstract and introduction would benefit from a brief statement of the specific UAV dynamics model (e.g., quadrotor equations) and the range of confined-space geometries used in the CFD campaign.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's significance. We address each major comment point by point below, with revisions made where the manuscript can be strengthened without altering its core contributions.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claims of 'comparable or exceeding' accuracy to TCN/MLP baselines and 'dramatic speed gains' (training from hours to milliseconds, inference to nanoseconds) are stated without any reported quantitative metrics, RMSE values, error bars, training/validation curves, or details on CFD data partitioning and cross-validation, leaving the performance assertions unsupported by evidence in the manuscript.

    Authors: The abstract serves as a concise summary, while the full manuscript provides the supporting quantitative evidence in the Results section, including RMSE comparisons to TCN/MLP baselines, error bars from repeated simulations, training/validation performance curves, and explicit details on CFD data partitioning (train/test split and cross-validation procedure). To directly address the concern, we have revised the abstract to incorporate key representative metrics and explicit references to the detailed evidence in the main text, ensuring the claims are now anchored by numbers without exceeding length constraints. revision: yes

  2. Referee: [PRC hardware model section] PRC hardware model section: no analysis or Monte-Carlo simulation is provided to quantify how laser intensity noise, phase drift, or fabrication tolerances in the optical feedback loop alter the reservoir state trajectory, eigenvalue spectrum, or prediction accuracy relative to the ideal numerical model; because closed-loop UAV stability results are obtained by feeding the simulated PRC outputs directly into the PID, any such degradation would directly undermine the claimed stability improvement and the assertion that millisecond training and nanosecond inference are preserved in flight.

    Authors: We agree that robustness to hardware non-idealities is essential for practical translation. The present work deliberately employs an idealized numerical model of the semiconductor laser with optical feedback to establish baseline performance and isolate the benefits of the photonic reservoir dynamics. In revision, we have added a dedicated paragraph in the PRC hardware model section that references published experimental characterizations of similar laser systems to estimate the effects of intensity noise and phase drift on state trajectories and readout accuracy; this shows that the core advantages (millisecond training, nanosecond inference) remain intact under realistic noise levels. A comprehensive Monte-Carlo study of fabrication tolerances is acknowledged as future experimental work and is now explicitly listed as such. The closed-loop UAV results are presented as model-based demonstrations, with the text clarified to avoid implying immediate hardware equivalence. revision: partial

Circularity Check

0 steps flagged

No significant circularity; central claims grounded in external CFD data and standard reservoir training.

full rationale

The derivation relies on CFD simulations of proximity flows as independent input data, with the deep PRC providing fixed reservoir dynamics from laser physics and only the linear readout trained via ridge regression on that external data. No equations reduce prediction accuracy, latency, or memory properties to a fitted parameter by construction, nor do any self-citations or ansatzes serve as load-bearing justifications for the core results. The comparison to TCN/MLP baselines and closed-loop UAV injection are demonstrated through simulation, keeping the performance claims falsifiable against the provided data rather than tautological.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the unverified assumption that CFD simulations faithfully reproduce real proximity flows and that photonic hardware noise will not degrade the linear readout performance; no free parameters beyond the standard ridge-regression hyperparameter are named, and no new physical entities are postulated.

free parameters (1)
  • ridge regression regularization parameter
    Standard hyperparameter for training the linear readout layer; value not reported in abstract.
axioms (1)
  • domain assumption High-fidelity CFD simulations accurately capture ground/ceiling effects and wake recirculation for the tested UAV geometries and speeds.
    Invoked when the abstract states that CFD data demonstrate the PRC accuracy.

pith-pipeline@v0.9.0 · 5533 in / 1419 out tokens · 47500 ms · 2026-05-10T15:38:15.710179+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages

  1. [1]

    Slope inspection under dense vegetation using lidar-based quadrotors,

    W. Liu, Y . Ren, R. Guo, V . W. W. Kong, A. S. P. Hung, F. Zhu, Y . Cai, H. Wu, Y . Zou, and F. Zhang, “Slope inspection under dense vegetation using lidar-based quadrotors,”Nature Communications, vol. 16, no. 1, p. 7411, 2025

    work page 2025

  2. [2]

    Drone-based displacement measurement of infrastructures utilizing phase information,

    S. Ri, J. Ye, N. Toyama, and N. Ogura, “Drone-based displacement measurement of infrastructures utilizing phase information,”Nature Communications, vol. 15, no. 1, p. 395, 2024

    work page 2024

  3. [3]

    Testing the delivery of human organ transportation with drones in the real world,

    A. T. Sage, M. Cypel, M. Cardinal, J. Qiu, A. Humar, and S. Keshav- jee, “Testing the delivery of human organ transportation with drones in the real world,”Science Robotics, vol. 7, no. 73, p. eadf5798, 2022

    work page 2022

  4. [4]

    Toward a fully autonomous uav: research platform for indoor and outdoor urban search and rescue,

    T. Tomic, K. Schmid, P. Lutz, A. Domel, M. Kassecker, E. Mair, I. L. Grixa, F. Ruess, M. Suppa, and D. Burschka, “Toward a fully autonomous uav: research platform for indoor and outdoor urban search and rescue,”IEEE Robotics & Automation Magazine, vol. 19, no. 3, pp. 46–56, 2012

    work page 2012

  5. [5]

    Modeling and control of a quadrotor uav with tilting propellers,

    M. Ryll, H. H. B ¨ulthoff, and P. R. Giordano, “Modeling and control of a quadrotor uav with tilting propellers,” in2012 IEEE International Conference on Robotics and Automation, 2012, pp. 4606–4613

    work page 2012

  6. [6]

    The effect of the ground on a helicopter rotor in forward flight,

    I. C. Cheeseman and W. E. Bennett, “The effect of the ground on a helicopter rotor in forward flight,” Aeronautical Research Council, Ministry of Supply, Tech. Rep. 3021, 1955

    work page 1955

  7. [7]

    Characterization of the aerodynamic ground effect and its influence in multirotor control,

    P. Sanchez-Cuevas, G. Heredia, and A. Ollero, “Characterization of the aerodynamic ground effect and its influence in multirotor control,” International Journal of Aerospace Engineering, vol. 2017, no. 1, p. 1823056, 2017

    work page 2017

  8. [8]

    Ceiling effects for surface locomotion of small rotorcraft,

    Y . H. Hsiao and P. Chirarattananon, “Ceiling effects for surface locomotion of small rotorcraft,” in2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 6214– 6219

    work page 2018

  9. [9]

    Neural lander: stable drone landing control using learned dynamics,

    G. Shi, X. Shi, M. O’Connell, R. Yu, K. Azizzadenesheli, A. Anand- kumar, Y . Yue, and S.-J. Chung, “Neural lander: stable drone landing control using learned dynamics,” in2019 International Conference on Robotics and Automation (ICRA), 2019, pp. 9784–9790

    work page 2019

  10. [10]

    Neural-fly enables rapid learning for agile flight in strong winds,

    M. O’Connell, G. Shi, X. Shi, K. Azizzadenesheli, A. Anandkumar, Y . Yue, and S.-J. Chung, “Neural-fly enables rapid learning for agile flight in strong winds,”Science Robotics, vol. 7, no. 66, p. eabm6597, 2022

    work page 2022

  11. [11]

    Neurobem: hybrid aerodynamic quadrotor model,

    L. Bauersfeld, E. Kaufmann, P. Foehn, S. Sun, and D. Scaramuzza, “Neurobem: hybrid aerodynamic quadrotor model,” inRobotics: Sci- ence and Systems XVII, 2021

    work page 2021

  12. [12]

    Data-driven modeling of ground effect for uav landing on a vertical oscillating platform,

    B. He, H. Zhang, B. Lai, S. Liu, and Y . Wang, “Data-driven modeling of ground effect for uav landing on a vertical oscillating platform,” in2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2024, pp. 8169–8174

    work page 2024

  13. [13]

    Neural- swarm2: planning and control of heterogeneous multirotor swarms using learned interactions,

    G. Shi, W. H ¨onig, X. Shi, Y . Yue, and S.-J. Chung, “Neural- swarm2: planning and control of heterogeneous multirotor swarms using learned interactions,”IEEE Transactions on Robotics, vol. 38, no. 2, pp. 1063–1079, 2022

    work page 2022

  14. [14]

    Compensating aerodynamics of over-actuated multi-rotor aerial platform with data-driven iterative learning control,

    P. Yu, Y . Su, L. Ruan, and T.-C. Tsao, “Compensating aerodynamics of over-actuated multi-rotor aerial platform with data-driven iterative learning control,”IEEE Robotics and Automation Letters, vol. 8, no. 10, pp. 6187–6194, 2023

    work page 2023

  15. [15]

    Neural-swarm: decen- tralized close-proximity multirotor control using learned interactions,

    G. Shi, W. H ¨onig, Y . Yue, and S.-J. Chung, “Neural-swarm: decen- tralized close-proximity multirotor control using learned interactions,” in2020 IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 3241–3247

    work page 2020

  16. [16]

    Nonlinear mpc for quadrotors in close-proximity flight with neural network downwash prediction,

    J. Li, L. Han, H. Yu, Y . Lin, Q. Li, and Z. Ren, “Nonlinear mpc for quadrotors in close-proximity flight with neural network downwash prediction,” in2023 62nd IEEE Conference on Decision and Control (CDC), 2023, pp. 2122–2128

    work page 2023

  17. [17]

    Learning quadrotor dynamics using neural network for flight control,

    S. Bansal, A. K. Akametalu, F. J. Jiang, F. Laine, and C. J. Tomlin, “Learning quadrotor dynamics using neural network for flight control,” in2016 IEEE 55th Conference on Decision and Control (CDC), 2016, pp. 4653–4660

    work page 2016

  18. [18]

    Physics-inspired temporal learn- ing of quadrotor dynamics for accurate model predictive trajectory tracking,

    A. Saviolo, G. Li, and G. Loianno, “Physics-inspired temporal learn- ing of quadrotor dynamics for accurate model predictive trajectory tracking,”IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 10 256–10 263, 2022

    work page 2022

  19. [19]

    Control of a tail-sitter vtol uav based on recurrent neural networks,

    J. Zhou, H. Xu, Z. Li, S. Shen, and F. Zhang, “Control of a tail-sitter vtol uav based on recurrent neural networks,”ArXiv, 2021. [Online]. Available: https://arxiv.org/abs/2104.02108

    work page arXiv 2021

  20. [20]

    Deep iterative learning control for quadrotor’s trajectory tracking,

    Z. Chen, X. Liang, and M. Zheng, “Deep iterative learning control for quadrotor’s trajectory tracking,” in2021 American Control Conference (ACC), 2021, pp. 1408–1413

    work page 2021

  21. [21]

    A gru network framework towards fault-tolerant control for flight vehicles based on a gain-scheduled approach,

    B. Yang, P. Lu, C. Du, and F. Cao, “A gru network framework towards fault-tolerant control for flight vehicles based on a gain-scheduled approach,”Aerospace Science and Technology, vol. 146, p. 108954, 2024

    work page 2024

  22. [22]

    On the difficulty of training recurrent neural networks,

    R. Pascanu, T. Mikolov, and Y . Bengio, “On the difficulty of training recurrent neural networks,” inProceedings of the 30th International Conference on Machine Learning, vol. 28, no. 3, 2013, pp. 1310–1318

    work page 2013

  23. [23]

    Parallel time-delay reservoir computing with quantum dot lasers,

    J. Y . Tang, B. D. Lin, J. Yu, X. He, and C. Wang, “Parallel time-delay reservoir computing with quantum dot lasers,”Journal of Quantum Electronics, vol. 58, no. 2, p. 8100109, 2022

    work page 2022

  24. [24]

    Deep photonic reservoir computer for nonlinear equalization of 16- level quadrature amplitude modulation signals,

    R. Q. Li, Y . W. Shen, Z. Niu, G. Xu, J. Yu, X. He, L. Yi, and C. Wang, “Deep photonic reservoir computer for nonlinear equalization of 16- level quadrature amplitude modulation signals,”APL Machine Learn- ing, vol. 3, p. 026113, 2025

    work page 2025

  25. [25]

    Parallel photonic information processing at gigabyte per second data rates using transient states,

    D. Brunner, M. C. Soriano, C. R. Mirasso, and I. Fischer, “Parallel photonic information processing at gigabyte per second data rates using transient states,”Nature Communications, vol. 4, p. 1364, 2013

    work page 2013

  26. [26]

    Brunner, M

    D. Brunner, M. C. Soriano, and G. Van der Sande,Photonic Reservoir Computing. Berlin, Germany: De Gruyter, 2019

    work page 2019

  27. [27]

    Deep photonic reservoir computing recurrent network,

    Y . W. Shen, R. Q. Li, G. T. Liu, J. Yu, X. He, L. Yi, and C. Wang, “Deep photonic reservoir computing recurrent network,” Optica, vol. 10, pp. 1745–1751, 2023

    work page 2023

  28. [28]

    Nakajima and I

    K. Nakajima and I. Fisher,Reservoir Computing: Theory, Physical Implementations, and Applications. Singapore: Springer, 2021

    work page 2021

  29. [29]

    Exper- imental demonstration of reservoir computing on a silicon photonics chip,

    K. Vandoorne, P. Mechet, T. Van Vaerenbergh, M. Fiers, G. Morthier, D. Verstraeten, B. Schrauwen, J. Dambre, and P. Bienstman, “Exper- imental demonstration of reservoir computing on a silicon photonics chip,”Nature Communications, vol. 5, p. 3541, 2014

    work page 2014

  30. [30]

    Information processing using a single dynamical node as complex system,

    L. Appeltant, M. C. Soriano, G. V . der Sande, J. Danckaert, S. Massar, J. Dambre, B. Schrauwen, C. R. Mirasso, and I. Fischer, “Information processing using a single dynamical node as complex system,”Nature Communications, vol. 2, no. 13, p. 468, 2011

    work page 2011

  31. [31]

    Asynchronous photonic time-delay reservoir computing,

    J. Y . Tang, B. D. Lin, Y . W. Shen, R. Q. Li, J. Yu, X. He, and C. Wang, “Asynchronous photonic time-delay reservoir computing,” Optics Express, vol. 31, pp. 2456–2466, 2023

    work page 2023

  32. [32]

    Role of delay-times in delay-based photonic reservoir computing,

    T. H ¨ulser, F. K¨oster, L. Jaurigue, and K. L¨udge, “Role of delay-times in delay-based photonic reservoir computing,”Optical Materials Express, vol. 12, no. 3, pp. 1214–1231, 2022

    work page 2022

  33. [33]

    Scalable wavelength-multiplexing photonic reservoir computing,

    R. Q. Li, Y . W. Shen, B. D. Lin, J. Yu, X. He, and C. Wang, “Scalable wavelength-multiplexing photonic reservoir computing,”APL Machine Learning, vol. 1, p. 036105, 2023

    work page 2023

  34. [34]

    A highly-efficient hybrid simulation system for flight controller design and evaluation of unmanned aerial vehicles,

    J. Wang, W. Song, Y . Fan, Y . Wang, and X. Liu, “A highly-efficient hybrid simulation system for flight controller design and evaluation of unmanned aerial vehicles,”ACM Trans. Graph., vol. 44, no. 6, Dec. 2025

    work page 2025

  35. [35]

    J. D. Anderson,Fundamentals of aerodynamics. New York: McGraw- Hill Education, 2010

    work page 2010