Energy-Efficient Autonomous Aerial Navigation with Dynamic Vision Sensors: A Physics-Guided Neuromorphic Approach

Amogh Joshi; Kaushik Roy; Manish Nagaraj; Rohan Kumar Manna; Sourav Sanyal

REVIEW 4 major objections 4 minor 20 references

Energy-Efficient Autonomous Aerial Navigation with Dynamic Vision Sensors: A Physics-Guided Neuromorphic Approach

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A physics-guided neuromorphic navigation system using event cameras and a spiking network cuts average drone flight time by about 20% and path length by around 15% in simulation.

desk verdict A plausible neuromorphic navigation proof-of-concept whose headline energy-efficiency claim is circular and unverified; worth a major-revision review, not a pass. read the letter →

arxiv 2502.05938 v2 pith:HMRNVAXK submitted 2025-02-09 cs.RO

Sourav Sanyal , Amogh Joshi , Manish Nagaraj , Rohan Kumar Manna , Kaushik Roy This is my paper · ORCID

classification cs.RO

keywords dynamicvisionsensoreventcameraspikingneuralnetworkphysics-guidedenergy-efficientnavigationautonomousdronemovinggatetrackingneuromorphic

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 claims that a drone can be steered through moving gates more efficiently by splitting perception from planning: an event camera plus a tiny spiking network tracks the gate asynchronously, while a physics-guided neural network maps depth to the energy-optimal flight time. The PgNN is trained on polynomial fits of a motor-energy model, so its predicted velocity encodes the trade-off between flight time and power consumption. In Gazebo simulation, the fused event-plus-depth system shortens average flight time by about 20% and path length by around 15% compared with a depth-only approach. The larger point is that neuromorphic sensing and physics-based planning together form a low-latency pipeline suitable for energy-constrained aerial robots.

What carries the argument

The load-bearing component is the energy-aware physics-guided neural network (PgNN), a three-layer fully connected network with hidden sizes [64, 128, 128], trained to predict the optimal velocity $v_{pred}$ from a depth input. Its training targets come from polynomial fits of an energy model $E(T)=\int_0^T \sum_j e_j(\tau)i_j(\tau)\,d\tau$ and a power law $P(t)=\kappa\|F_{\rm thrust}(t)\|^\alpha$; the optimal velocity $v_{\rm opt}$ is the zero of $dE/dv$, and the predicted flight time is $t_{\rm traj}=d/v_{\rm pred}$. A composite loss $L_{PgNN}=L_{data}+\lambda_1 L_{physics}+\lambda_2 L_{energy}$ ties the predictions to those physics constraints. Around it, a single-layer LIF spiking network with a $3\times3$ kernel detects and bounds the gate from event streams, and a rule-based planner computes the gate's future position $y^*=y_2+v_r t_{\rm traj}$ (with a bounce correction) so the drone arrives as the gate swings past.

What would settle it

Run the same moving-gate scenario on a physical quadrotor with both the event-plus-PgNN controller and the depth-only controller, recording battery energy with coulomb counting and flight time; if the real energy-velocity curve lacks a minimum at the predicted $v_{\rm opt}$, or the measured savings do not approach 20% and 15%, the central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that the energy-optimal time to cross a gate can be predicted from depth alone, while event data handles the perception, and that this division makes navigation both faster and cheaper. The paper's simulations show the neuromorphic approach producing shorter paths at every depth tested, for example a path length of 3.5 m versus 4.6 m at 2 m depth for an off-center start, and an average reduction of roughly 20% in flight time and 15% in path length versus the depth-based method. The result is presented as evidence that event-based vision and physics-guided planning can be integrated into a neurosymbolic navigation stack for dynamic obstacle environments.

Load-bearing premise

The claimed energy savings are computed with the same energy model (Eqs. 8 and 11) that produced the PgNN's training targets, so if that model does not match real motor and aerodynamic behavior, the savings may not appear on hardware.

Editorial extensions

If this is right

The event-based detector keeps tracking a moving gate from 2 m to 9 m depth using only nine spiking neurons and a single $3\times3$ convolution, with mean IoU between roughly 0.60 and 0.83.
Fusing event and depth data shortens paths compared with depth-only perception, including 3.5 m versus 4.6 m at 2 m depth and 5.9 m versus 7.6 m at 5 m depth for off-center starts.
The PgNN's predicted flight time lets the symbolic planner anticipate the gate's bounce and velocity, so the drone can pass through the moving ring rather than chase its current position.
Averaged over the tested starting positions and depths, the neuromorphic approach reduces flight time by about 20% and path length by about 15% relative to the depth-based method.
The energy-optimal velocity depends on depth and on the power exponent $\alpha$, so the framework is portable to another vehicle only after recalibrating $\alpha$ and the regularization weights.

Reading between the lines

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

If the same energy model holds on hardware, the depth-to-velocity mapping should transfer to other multirotors after recalibrating $\alpha$ and $\lambda$, because the PgNN input is just depth and the SNN needs no labels.
The event camera's advantage should grow as the gate moves faster, since frame-based sensing degrades with motion blur exactly where events become denser; sweeping gate speed while measuring tracking IoU would test this.
A decisive test of the energy claim would measure battery energy with coulomb counting on a physical drone, using an independently calibrated motor model, so the training model and evaluation model are not the same.
The rule-based gate-passing planner is a low-level reactive primitive that could be composed with higher-level race or waypoint planners without modifying either neural module.

Share X Bluesky LinkedIn Reddit HN

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

Short version: this is a plausible integration demo from the EV-Planner line. The moving-gate testbed and the alpha/lambda sensitivity sweep are new, and the system is clearly built and demonstrated in simulation. The headline energy claim, however, is not established: the same energy model supplies both the PgNN training targets and the evaluation metric, so the 'near-minimum-energy' result is a self-consistency check, not a validation.

What is genuinely good: the SNN front end is extremely light (one 3x3 convolution, 9 spiking neurons, unsupervised), and the IoU-vs-depth curves show it tracks a moving gate out to around 6 m with mIOU 0.6-0.65 at the far end. The PgNN is a straightforward 3-layer MLP and the moving-gate predictor is a clean rule-based module. The ROS/Gazebo implementation looks real; the trajectory plots in Figure 8 are legible and the system does navigate through a moving gate. That is a legitimate proof-of-concept.

The soft spots are in the evaluation. First, the energy-optimality claim is circular. The polynomial fits in Section III-B produce v_opt from energy-velocity data generated by Equations (8) and (11), and Section IV computes 'dynamic energy' with the same equations. Nothing independent checks that model. Calling alpha=0.2 'tuned to our Bebop simulation' is not calibration; no identification error, no power trace, no error bars. If the model is wrong, the predicted v_opt is not energy-minimal and the evaluation cannot detect it. This does not invalidate the integration, but it invalidates the energy claim as stated.

Second, the depth-based baseline is unspecified. Figure 8 shows red trajectories, but no details on the planner, controller, or sensor setup. The 20%/15% flight-time and path-length improvements are therefore unverifiable. There are no repeated trials or error bars. Code and data are not released. These are fixable, but they need to be fixed.

The citation pattern is honest about prior work; most components come from the same group's earlier papers, which is incremental but not deceptive.

Who is this for? Someone tracking neuromorphic perception and planning demos will find it a useful data point. Someone wanting evidence that DVS plus PgNN saves real energy will not. I would send it to peer review with a request for major revision: specify the baseline, add repeated trials, validate energy against an independent model or hardware, and release code. I would not cite it for the energy numbers.

Referee Report

4 major / 4 minor

Summary. The paper proposes a neuromorphic navigation framework for a quadrotor passing through moving gates, integrating an event camera with a shallow spiking neural network (SNN) for object detection, a physics-guided neural network (PgNN) for predicting flight velocity/time, and a rule-based planner for gate interception. The system is implemented in the Gazebo simulator with ROS. The central claims are (i) the PgNN produces near-minimum-energy trajectories and (ii) the neuromorphic approach reduces average flight time by approximately 20% and path length by around 15% relative to a depth-based method.

Significance. If the core claims were substantiated, the work would be a useful demonstration of combining event-based perception with physics-informed planning for energy-efficient aerial navigation. The SNN's minimal parameter count and the integration of a physics-guided network with a rule-based planner are conceptually appealing and align with current interests in neuromorphic low-latency autonomy. However, the paper as presented does not provide the evidence needed to support its headline claims: the energy-optimality result is not independently verified, and the comparison baseline is not specified. The strengths of the paper are its clear system concept and the inclusion of a sensitivity analysis over the power-model exponent and regularization weights, but these are not sufficient to establish the claimed benefits.

major comments (4)

[Section IV.C] The claim that the physics-guided neuromorphic approach reduces average flight time by approximately 20% and path length by around 15% compared to the depth-based method is not verifiable as stated. The 'depth-based method' (red trajectory in Figure 8) is never defined: no algorithm, parameters, perception pipeline, or planner configuration are given. Without this baseline specification, the comparison cannot be reproduced or attributed to the neuromorphic versus frame-based processing distinction. Moreover, no error bars, standard deviations, or number of repeated trials are reported, so the 20%/15% figures may be within run-to-run variability.
[Sections III.B and IV.B] The 'near-minimum energy' claim is circular and not directly tested. The PgNN training targets (v_opt in Table I) are obtained by fitting 5th-degree polynomials to energy-velocity data generated by the same energy model (Eqs. 8 and 11) that is later used to compute the reported 'dynamic energy' in Figure 7. No independent power model, hardware trace, or alternative simulation is used to validate the energy model. Furthermore, the paper never compares the PgNN's achieved energy to the theoretical minimum E(v_opt) from Eq. (9); Figure 7 only shows how energy varies with α and λ. In addition, the note in Section IV.B admits that the training data come from straight-line flights while the evaluation flights have longer curved paths; the optimal velocity for a straight-line segment is not necessarily optimal for the obstacle-avoidance trajectories actually flown, so even within the model the PgNN's velocities may not be near-minimum for the deployed task.
[Section III.B and Table I] The derivation of the optimal velocity v_opt is not reproducible. The paper does not report the polynomial coefficients, the range of velocities and depths used, the number of data points, or the goodness of fit for the 5th-degree polynomial. Table I is only a symbolic template with ellipses, not actual numerical samples. Without these details, it is impossible to verify that the derivative condition in Eq. (9) yields a meaningful global minimum or to reconstruct the training set for the PgNN.
[Section III.C, Eq. (14)] The physics consistency loss L_physics is underspecified. The term ||x_pred - x_sim|| relies on states x_sim from 'a physics-based simulation model (Equation 3)', but Eq. (3) is a generic quadrotor model with an unspecified drag function d(v). The paper does not explain how x_sim is computed, whether x_pred is a position or velocity prediction, or how this loss is evaluated (e.g., over what time horizon). As a load-bearing component of the PgNN training objective, this needs a concrete definition.

minor comments (4)

[Section III.A, Eq. (4)] The LIF neuron update equation appears garbled: V[t] = β, V[t_{n-1}] + W, X[t] is not a valid mathematical expression and likely should be V[t] = β V[t_{n-1}] + W X[t].
[Throughout] There are several typos and formatting issues: 'minimaly' in the bullet list, 'FIITING' in the Table I caption, 'UA V' instead of 'UAV' in Section III.B, and the floor-function notation in Eq. (6) is mis-rendered as 'j Xmax − Xmin 2 k'.
[Section III.A] The SNN is described as 'unsupervised', but the paragraph states that β = 0.1 and Vth = 1.75 were obtained by fine-tuning for the specific gate moving at 4 m/s. Calling this 'unsupervised' is misleading; it requires per-scenario threshold tuning.
[Section II.B] The sentence 'A spike is generated whenever V (t) exceeds a threshold Vth, at which point V (t) is reset to Vreset' uses a strange arrow notation '− − − − − − →' that is not defined.

Circularity Check

1 steps flagged · score 6.0 of 10

Energy-optimality claim is self-referential: PgNN training targets and the Section IV energy evaluation both come from the same Eq. (8)/(11) power model, so 'near-minimum energy' is not independently validated.

fitted input called prediction [Section III.B, Eqs. (8)-(11), Table I; Section IV.B, Fig. 7]
"By fitting a 5th-degree polynomial to the energy-velocity data, the PgNN captures the non-linear dynamics inherent in the system. The optimal velocity vopt is determined by finding the velocity at which the derivative of the energy function equals zero: dE(v)/dv = 0 ⇒ vopt = arg min_v E(v). ... P(t) = κ∥Fthrust(t)∥α. ... The samples used to train the PgNN are summarized in Table I, which lists various depths dn, their corresponding optimal velocities vn, and the associated constraints derived from the polynomial fits of E(v) and its derivative. ..."

The PgNN's regression target vopt is generated by differentiating the same energy model that later scores the system: Eq. (8) integrates motor electrical energy and Eq. (11) sets P = κ||F_thrust||^α; a 5th-degree polynomial fit of E(v) yields vopt, and Table I feeds these vopt to the PgNN as 'ground truth'. In Section IV.B, the reported 'dynamic energy' in Fig. 7 is computed with the same Eq. (11) power law, with α swept as a sensitivity parameter. Hence the network is trained and evaluated on the same objective function: the 'near-minimum energy' result cannot detect any error in the energy model itself, and the prediction is essentially a consistency check on the polynomial fit. No independent power trace, hardware measurement, or external energy benchmark is provided.

full rationale

The central energy-efficiency claim is partially circular. The PgNN's training labels (vopt per depth) are produced by 5th-degree polynomial fits to E(v) assembled from the motor model Eq. (8) and the power law Eq. (11), and Table I explicitly states these samples and constraints come from those fits. The evaluation in Section IV.B then reports 'dynamic energy' using the same Eq. (11) model (with α treated as a tuning parameter), so the 'near-minimum energy' assertion is a self-consistency check rather than a test against independent ground truth. The physics-consistency loss Eq. (14) also compares against the same simulation model Eq. (3), which makes training self-consistent but does not validate the energy model externally. The α = 0.2 'tuned to our Bebop simulation' is a stated assumption, not a calibrated identification with error bars or a hardware power trace. The separate 20%/15% flight-time and path-length comparison against the depth-based baseline in Section IV.C is a genuine trajectory-level comparison, but the depth-based baseline is never specified, so reproducibility is limited; this is a methodology concern rather than circularity. No load-bearing self-citation chain was found: prior work [14], [16], [17] is used for architecture and detection ideas, but the circularity resides in fitting and evaluating with the same energy equations, not in the citations.

Assumptions & free parameters 8 free parameters · 6 assumptions · 0 invented entities

The central contribution rests on assumed energy and power models, the fidelity of the Gazebo event-camera simulation, and the correctness of polynomial fits. No new physical entity is introduced.

free parameters (8)

SNN leak factor beta = 0.1
Hand-tuned for the specific gate moving at 4 m/s (Section III.A).
SNN firing threshold Vth = 1.75
Fine-tuned for the given gate speed 4 m/s (Section III.A).
SNN kernel W = 3x3 kernel, weights unspecified
Described as learnable but no training procedure is given; effectively a hand-chosen or randomly initialized filter.
Event camera contrast threshold C = not specified
User-defined in Eq. 1; value is never reported.
Power model exponent alpha = 0.2, 0.3, 0.5
0.2 is tuned to the Bebop simulation; 0.3 and 0.5 are hypothetical (Section IV.B).
Regularization weights lambda1, lambda2 = not specified
Tuned to balance loss terms in Eq. 12; exact values not listed.
Polynomial coefficients per depth = 6 coefficients per depth (5th degree)
Fitted to simulated energy-velocity data for each depth (Section III.B, Table I).
Gate speed for SNN tuning = 4 m/s
The SNN thresholds are tuned for this speed; generalization to other speeds is not tested.

assumptions (6)

domain assumption Energy model E(T) = integral of sum e_j i_j over time (Eq. 8) accurately represents quadrotor energy consumption.
The entire optimality argument uses this model; no real-world validation is provided.
domain assumption Power model P = kappa * ||F_thrust||^alpha (Eq. 11) is valid, with kappa and alpha constants.
Introduced as a given actuator model; alpha is varied but no independent calibration is shown except for 0.2 labeled as tuned to a Bebop simulation.
ad hoc to paper The 5th-degree polynomial fit of E(v) is a faithful approximation of the energy-velocity relationship.
The polynomial is fit to simulation data for each depth and then used to define v_opt via its derivative; the fit quality is not shown.
standard math v_opt is the minimizer of E(v) and satisfies dE/dv = 0.
Calculus minimum condition, but applied to the fitted polynomial, not a proven physical optimum.
domain assumption Event rate is proportional to object speed (Eq. 5).
A heuristic used to justify the SNN's ability to detect fast-moving gates; no empirical evidence is provided within the paper.
domain assumption The Gazebo simulator and its event camera plugin faithfully reproduce DVS events and drone dynamics.
The entire experimental evaluation relies on this simulation fidelity; no real-world comparison is made.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Energy-Efficient Autonomous Aerial Navigation with Dynamic Vision Sensors: A Physics-Guided Neuromorphic Approach." pith.science (2026). https://pith.science/paper/HMRNVAXK

@misc{pith2026250205938,
  author       = {Pith},
  title        = {Pith review of: Energy-Efficient Autonomous Aerial Navigation with Dynamic Vision Sensors: A Physics-Guided Neuromorphic Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HMRNVAXK}},
  note         = {Machine review of arXiv:2502.05938}
}

read the original abstract

Vision-based object tracking is a critical component for achieving autonomous aerial navigation, particularly for obstacle avoidance. Neuromorphic Dynamic Vision Sensors (DVS) or event cameras, inspired by biological vision, offer a promising alternative to conventional frame-based cameras. These cameras can detect changes in intensity asynchronously, even in challenging lighting conditions, with a high dynamic range and resistance to motion blur. Spiking neural networks (SNNs) are increasingly used to process these event-based signals efficiently and asynchronously. Meanwhile, physics-based artificial intelligence (AI) provides a means to incorporate system-level knowledge into neural networks via physical modeling. This enhances robustness, energy efficiency, and provides symbolic explainability. In this work, we present a neuromorphic navigation framework for autonomous drone navigation. The focus is on detecting and navigating through moving gates while avoiding collisions. We use event cameras for detecting moving objects through a shallow SNN architecture in an unsupervised manner. This is combined with a lightweight energy-aware physics-guided neural network (PgNN) trained with depth inputs to predict optimal flight times, generating near-minimum energy paths. The system is implemented in the Gazebo simulator and integrates a sensor-fused vision-to-planning neuro-symbolic framework built with the Robot Operating System (ROS) middleware. This work highlights the future potential of integrating event-based vision with physics-guided planning for energy-efficient autonomous navigation, particularly for low-latency decision-making.

Figures

Figures reproduced from arXiv: 2502.05938 by the authors.

**Figure 1.** Event-Based Vision and Spiking Neuron Model: Illustration of the conceptual flow from event-based sensing to biological inspiration and spiking neural networks (SNNs). Left: An event camera outputs discrete intensity changes over time (red/blue dots) rather than continuous image frames. Center: Biological neurons communicate via discrete spikes, with dendrites receiving inputs that the soma integrates before generat… view at source ↗

**Figure 2.** System architecture: Integrating neuromorphic vision, physics-guided neural networks, and symbolic rule-based reasoning. Event and depth streams feed the SNN and PgNN, informing real-time navigation decisions. Adapted from [14]. The motion planner is based on the work in [15]. C. Physics-Guided Neural Networks (PgNNs) Physics-Guided Neural Networks (PgNNs) [9]–[13] integrate system dynamics or physical constraints … view at source ↗

**Figure 3.** Event-based Object Detection at Various Depths. Each sub-panel shows neuromorphic event output and a bounding box around a moving gate. Although event density decreases with increasing depth, the SNN continues to isolate and track the gate in real time. that scales the contribution of the input, and X[t] is the aggregated spike input at time t. The leak factor β controls how quickly the neuron forgets previous input… view at source ↗

Figures from the paper (6 more)

**Figure 4.** Figure 4: Optimal velocity and energy-consumption patterns. Each depth [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Drone navigating through a moving gate in Gazebo simulation. The gate moves from right to left, while the drone starts at position [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: Tracking performance vs. depth. The plot compares the mean IoU [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

**Figure 7.** Figure 7: Comparison of flight time, path length, and dynamic energy across [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

**Figure 8.** Figure 8: Comparison of Navigation Trajectories: The figure illustrates the navigation paths of a drone using depth-based perception (red trajectory) versus a [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

**Figure 9.** Figure 9: Flight time (left y-axis) and path length (right-y axis) as functions [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 15 canonical work pages

[1]

A low power, fully event-based gesture recognition system,

A. Amir, B. Taba, D. Berg, T. Melano, J. McKinstry, C. Di Nolfo, T. Nayak, A. Andreopoulos, G. Garreau, M. Mendoza et al. , “A low power, fully event-based gesture recognition system,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 7243–7252

2017
[2]

A 128 ×128 120 db 15 µs latency asynchronous temporal contrast vision sensor,

P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128 ×128 120 db 15 µs latency asynchronous temporal contrast vision sensor,” IEEE journal of solid-state circuits , vol. 43, no. 2, pp. 566–576, 2008

work page 2008
[3]

A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds,

C. Posch, D. Matolin, and R. Wohlgenannt, “A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds,” IEEE Journal of Solid-State Cir- cuits, vol. 46, no. 1, pp. 259–275, 2010

work page 2010
[4]

Spikenet: A simulator for modeling large networks of integrate and fire neurons,

A. Delorme, J. Gautrais, R. Van Rullen, and S. Thorpe, “Spikenet: A simulator for modeling large networks of integrate and fire neurons,” Neurocomputing, vol. 26, pp. 989–996, 1999

work page 1999
[5]

Spike-flownet: event-based optical flow estimation with energy-efficient hybrid neural networks,

C. Lee, A. K. Kosta, A. Z. Zhu, K. Chaney, K. Daniilidis, and K. Roy, “Spike-flownet: event-based optical flow estimation with energy-efficient hybrid neural networks,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, 2020, pp. 366–382

work page 2020
[6]

Spiking neural network-based control of an unmanned aerial system implemented on a customized neural flight simulation environment,

O. A. Garcia A., D. C. Arana, E. S. Espinoza, I. R. Scola, L. R. Garcia Carrillo, and A. T. Sornborger, “Spiking neural network-based control of an unmanned aerial system implemented on a customized neural flight simulation environment,” in 2024 American Control Conference (ACC), 2024, pp. 3124–3129

work page 2024
[7]

Predicting visually-modulated precisely-timed spikes across a coordinated and comprehensive motor program,

U. B. Sikandar, H. Choi, J. Putney, H. Yang, S. Ferrari, and S. Spon- berg, “Predicting visually-modulated precisely-timed spikes across a coordinated and comprehensive motor program,” in 2023 International Joint Conference on Neural Networks (IJCNN) , 2023, pp. 1–8

work page 2023
[8]

A relative spike-timing approach to kernel-based decoding demonstrated for insect flight experiments,

H. Yang, J. Putney, U. Bin Sikandar, P. Zhu, S. Sponberg, and S. Ferrari, “A relative spike-timing approach to kernel-based decoding demonstrated for insect flight experiments,” in 2022 International Joint Conference on Neural Networks (IJCNN) , 2022, pp. 1–7

work page 2022

Show all 20 references

[9]

Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,

M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics , vol. 378, pp. 686–707, 2019

2019
[10]

Physics-informed machine learning,

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang, “Physics-informed machine learning,” Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021

2021
[11]

Physics-informed neural networks-based model predictive control for multi-link manip- ulators,

J. Nicodemus, J. Kneifl, J. Fehr, and B. Unger, “Physics-informed neural networks-based model predictive control for multi-link manip- ulators,” IFAC-PapersOnLine, vol. 55, no. 20, pp. 331–336, 2022, 10th Vienna International Conference on Mathematical Modelling MATHMOD 2022

2022
[12]

Ramp-net: A robust adaptive mpc for quadrotors via physics-informed neural network,

S. Sanyal and K. Roy, “Ramp-net: A robust adaptive mpc for quadrotors via physics-informed neural network,” in2023 International Conference on Robotics and Automation (ICRA) . IEEE, 2023, pp. 1019–1025

2023
[13]

Knode-mpc: A knowledge- based data-driven predictive control framework for aerial robots,

K. Y . Chee, T. Z. Jiahao, and M. A. Hsieh, “Knode-mpc: A knowledge- based data-driven predictive control framework for aerial robots,”IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 2819–2826, 2022

2022
[14]

Ev-planner: Energy-efficient robot navigation via event-based physics-guided neuromorphic plan- ner,

S. Sanyal, R. K. Manna, and K. Roy, “Ev-planner: Energy-efficient robot navigation via event-based physics-guided neuromorphic plan- ner,” IEEE Robot Autom Lett , vol. 9, no. 3, pp. 2080–2087, 2024

2024
[15]

Minimum snap trajectory generation and control for quadrotors,

D. Mellinger and V . Kumar, “Minimum snap trajectory generation and control for quadrotors,” in 2011 IEEE international conference on robotics and automation . IEEE, 2011, pp. 2520–2525

2011
[16]

Dotie - detecting ob- jects through temporal isolation of events using a spiking architecture,

M. Nagaraj, C. M. Liyanagedera, and K. Roy, “Dotie - detecting ob- jects through temporal isolation of events using a spiking architecture,” in 2023 International Conference on Robotics and Automation (ICRA). IEEE, 2023, pp. 4858–4864

2023
[17]

Real-time neuromorphic navigation: Integrating event-based vision and physics-driven planning on a parrot bebop2 quadrotor,

A. Joshi, S. Sanyal, and K. Roy, “Real-time neuromorphic navigation: Integrating event-based vision and physics-driven planning on a parrot bebop2 quadrotor,” 2024. [Online]. Available: https: //arxiv.org/abs/2407.00931

2024 arXiv
[18]

Object detection using yolo: Challenges, architectural successors, datasets and applications,

T. Diwan, G. Anirudh, and J. V . Tembhurne, “Object detection using yolo: Challenges, architectural successors, datasets and applications,” multimedia Tools and Applications , vol. 82, no. 6, pp. 9243–9275, 2023

2023
[19]

Faster r-cnn: Towards real-time object detection with region proposal networks,

S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-time object detection with region proposal networks,” IEEE transactions on pattern analysis and machine intelligence , vol. 39, no. 6, pp. 1137– 1149, 2016

2016
[20]

A survey of the hough transform,

J. Illingworth and J. Kittler, “A survey of the hough transform,” Computer vision, graphics, and image processing , vol. 44, no. 1, pp. 87–116, 1988

1988

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.

[1] [1]

A low power, fully event-based gesture recognition system,

A. Amir, B. Taba, D. Berg, T. Melano, J. McKinstry, C. Di Nolfo, T. Nayak, A. Andreopoulos, G. Garreau, M. Mendoza et al. , “A low power, fully event-based gesture recognition system,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 7243–7252

2017

[2] [2]

A 128 ×128 120 db 15 µs latency asynchronous temporal contrast vision sensor,

P. Lichtsteiner, C. Posch, and T. Delbruck, “A 128 ×128 120 db 15 µs latency asynchronous temporal contrast vision sensor,” IEEE journal of solid-state circuits , vol. 43, no. 2, pp. 566–576, 2008

work page 2008

[3] [3]

A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds,

C. Posch, D. Matolin, and R. Wohlgenannt, “A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds,” IEEE Journal of Solid-State Cir- cuits, vol. 46, no. 1, pp. 259–275, 2010

work page 2010

[4] [4]

Spikenet: A simulator for modeling large networks of integrate and fire neurons,

A. Delorme, J. Gautrais, R. Van Rullen, and S. Thorpe, “Spikenet: A simulator for modeling large networks of integrate and fire neurons,” Neurocomputing, vol. 26, pp. 989–996, 1999

work page 1999

[5] [5]

Spike-flownet: event-based optical flow estimation with energy-efficient hybrid neural networks,

C. Lee, A. K. Kosta, A. Z. Zhu, K. Chaney, K. Daniilidis, and K. Roy, “Spike-flownet: event-based optical flow estimation with energy-efficient hybrid neural networks,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, 2020, pp. 366–382

work page 2020

[6] [6]

Spiking neural network-based control of an unmanned aerial system implemented on a customized neural flight simulation environment,

O. A. Garcia A., D. C. Arana, E. S. Espinoza, I. R. Scola, L. R. Garcia Carrillo, and A. T. Sornborger, “Spiking neural network-based control of an unmanned aerial system implemented on a customized neural flight simulation environment,” in 2024 American Control Conference (ACC), 2024, pp. 3124–3129

work page 2024

[7] [7]

Predicting visually-modulated precisely-timed spikes across a coordinated and comprehensive motor program,

U. B. Sikandar, H. Choi, J. Putney, H. Yang, S. Ferrari, and S. Spon- berg, “Predicting visually-modulated precisely-timed spikes across a coordinated and comprehensive motor program,” in 2023 International Joint Conference on Neural Networks (IJCNN) , 2023, pp. 1–8

work page 2023

[8] [8]

A relative spike-timing approach to kernel-based decoding demonstrated for insect flight experiments,

H. Yang, J. Putney, U. Bin Sikandar, P. Zhu, S. Sponberg, and S. Ferrari, “A relative spike-timing approach to kernel-based decoding demonstrated for insect flight experiments,” in 2022 International Joint Conference on Neural Networks (IJCNN) , 2022, pp. 1–7

work page 2022