Publications Neural Network Control
Neuroflight: Next Generation Flight Control Firmware
, September 2019, Boston, MA, USA
Abstract
Quadcopter aircraft are now mainstream and used in a number of contexts, including surveillance, aerial mapping and surveying, natural disaster recovery, and search and rescue operations. In the last decade, seminal results have been accomplished in autonomous and semi-autonomous flight. Works in this domain have extensively addressed challenges in state estimation, trajectory tracking, localization, and the like. Such high-level control objectives rely on, and are directly impacted by, the behavior of the inner-most control loop, namely attitude control. Remarkably however, attitude control has received little-to-no attention in recent years. Flight control implementations predominantly use the classical PID algorithm. PID is easy to implement, which resonates well with resource constrained quadcopters. Nonetheless, PID has important limitations in terms of adaptability. Also, its performance is only as good as the tune. In this work, we investigate if performing better than PID is possible, while still abiding to the typical resource constraints of small-scale quadcopters. For this purpose, we introduce Neuroflight, the first open-source neural network-based flight controller firmware. We present our toolchain for training a neural network in simulation and compiling it to run on embedded hardware. We discuss the main challenges faced when jumping from simulation to reality along with the proposed solutions. Our evaluation shows that the neural network controller can execute at over 2kHz on an ARM Cortex-M7 processor. We demonstrate via flight tests that a quadcopter running Neuroflight can achieve stable flight and execute aerobatic maneuvers. arXiv:1901.06553v2 [cs.RO] 16 Sep 2019
Code and hardware artifacts
Cite this paper
@inproceedings{koch2019neuroflight,
title = {{Neuroflight: Next Generation Flight Control Firmware}},
author = {Koch, William and Mancuso, Renato and Bestavros, Azer},
year = 2019,
month = sep,
publisher = {Boston University (BU)},
address = {Boston, MA, USA},
url = {https://arxiv.org/pdf/1901.06553}
}
TY - CONF AU - Koch, William AU - Mancuso, Renato AU - Bestavros, Azer TI - Neuroflight: Next Generation Flight Control Firmware T2 - PY - 2019 DA - 2019/09// PB - Boston University (BU) CY - Boston, MA, USA AB - Quadcopter aircraft are now mainstream and used in a number of contexts, including surveillance, aerial mapping and surveying, natural disaster recovery, and search and rescue operations. In the last decade, seminal results have been accomplished in autonomous and semi-autonomous flight. Works in this domain have extensively addressed challenges in state estimation, trajectory tracking, localization, and the like. Such high-level control objectives rely on, and are directly impacted by, the behavior of the inner-most control loop, namely attitude control. Remarkably however, attitude control has received little-to-no attention in recent years. Flight control implementations predominantly use the classical PID algorithm. PID is easy to implement, which resonates well with resource constrained quadcopters. Nonetheless, PID has important limitations in terms of adaptability. Also, its performance is only as good as the tune. In this work, we investigate if performing better than PID is possible, while still abiding to the typical resource constraints of small-scale quadcopters. For this purpose, we introduce Neuroflight, the first open-source neural network-based flight controller firmware. We present our toolchain for training a neural network in simulation and compiling it to run on embedded hardware. We discuss the main challenges faced when jumping from simulation to reality along with the proposed solutions. Our evaluation shows that the neural network controller can execute at over 2kHz on an ARM Cortex-M7 processor. We demonstrate via flight tests that a quadcopter running Neuroflight can achieve stable flight and execute aerobatic maneuvers. arXiv:1901.06553v2 [cs.RO] 16 Sep 2019 ER -
%0 Conference Paper %A Koch, William %A Mancuso, Renato %A Bestavros, Azer %T Neuroflight: Next Generation Flight Control Firmware %B %D 2019 %I Boston University (BU) %C Boston, MA, USA %X Quadcopter aircraft are now mainstream and used in a number of contexts, including surveillance, aerial mapping and surveying, natural disaster recovery, and search and rescue operations. In the last decade, seminal results have been accomplished in autonomous and semi-autonomous flight. Works in this domain have extensively addressed challenges in state estimation, trajectory tracking, localization, and the like. Such high-level control objectives rely on, and are directly impacted by, the behavior of the inner-most control loop, namely attitude control. Remarkably however, attitude control has received little-to-no attention in recent years. Flight control implementations predominantly use the classical PID algorithm. PID is easy to implement, which resonates well with resource constrained quadcopters. Nonetheless, PID has important limitations in terms of adaptability. Also, its performance is only as good as the tune. In this work, we investigate if performing better than PID is possible, while still abiding to the typical resource constraints of small-scale quadcopters. For this purpose, we introduce Neuroflight, the first open-source neural network-based flight controller firmware. We present our toolchain for training a neural network in simulation and compiling it to run on embedded hardware. We discuss the main challenges faced when jumping from simulation to reality along with the proposed solutions. Our evaluation shows that the neural network controller can execute at over 2kHz on an ARM Cortex-M7 processor. We demonstrate via flight tests that a quadcopter running Neuroflight can achieve stable flight and execute aerobatic maneuvers. arXiv:1901.06553v2 [cs.RO] 16 Sep 2019
[
{
"id": "koch2019neuroflight",
"type": "paper-conference",
"title": "Neuroflight: Next Generation Flight Control Firmware",
"author": [
{
"family": "Koch",
"given": "William"
},
{
"family": "Mancuso",
"given": "Renato"
},
{
"family": "Bestavros",
"given": "Azer"
}
],
"container-title": "",
"issued": {
"date-parts": [
[
2019,
9
]
]
},
"publisher": "Boston University (BU)",
"publisher-place": "Boston, MA, USA",
"abstract": "Quadcopter aircraft are now mainstream and used in a number of contexts, including surveillance, aerial mapping and surveying, natural disaster recovery, and search and rescue operations. In the last decade, seminal results have been accomplished in autonomous and semi-autonomous flight. Works in this domain have extensively addressed challenges in state estimation, trajectory tracking, localization, and the like. Such high-level control objectives rely on, and are directly impacted by, the behavior of the inner-most control loop, namely attitude control. Remarkably however, attitude control has received little-to-no attention in recent years. Flight control implementations predominantly use the classical PID algorithm. PID is easy to implement, which resonates well with resource constrained quadcopters. Nonetheless, PID has important limitations in terms of adaptability. Also, its performance is only as good as the tune. In this work, we investigate if performing better than PID is possible, while still abiding to the typical resource constraints of small-scale quadcopters. For this purpose, we introduce Neuroflight, the first open-source neural network-based flight controller firmware. We present our toolchain for training a neural network in simulation and compiling it to run on embedded hardware. We discuss the main challenges faced when jumping from simulation to reality along with the proposed solutions. Our evaluation shows that the neural network controller can execute at over 2kHz on an ARM Cortex-M7 processor. We demonstrate via flight tests that a quadcopter running Neuroflight can achieve stable flight and execute aerobatic maneuvers. arXiv:1901.06553v2 [cs.RO] 16 Sep 2019"
}
]
W. Koch, R. Mancuso, and A. Bestavros, “Neuroflight: Next Generation Flight Control Firmware,” in , Sep. 2019.
Koch, W., Mancuso, R., & Bestavros, A. (2019). Neuroflight: Next Generation Flight Control Firmware. In . Boston University (BU).