While operating spacecraft, many teams break operations planning into two processes. One process focuses on planning for the upcoming cycle, while another process governs the development of unique activities that require careful study, testing, and review before they are ready for use in flight. While developing unique activities that require extra testing and review, teams often make use of testbeds on earth to ensure the activity is well planned and that sequences of commands are compatible with hardware and software configurations. Despite careful tracking, it can be easy for mistakes to creep in before sequence delivery and uplink while teams are testing and iterating over difference sequence versions. For example, a sequence that required an update for success in the testbed could have an earlier version inadvertently submitted to the flight sequence database. The Engineering Operations team for NASA's Mars Science Laboratory Curiosity Rover had a desire to improve the traceability of these sequences as they are tested, reviewed, and uplinked. This paper details the design, development, and implementation of a new tool that uses the Sequence Checksum to make it easier for operators to prove three key things during the review process. First, that the sequences evaluated in the testbed match what has been delivered to the sequence database for use in flight. Second, that all sequences being delivered as part of an activity were indeed tested. Third, that all sequences necessary for activity completion have been delivered to the sequence database. While this may be trivial to accomplish for simple activities, keeping track of all this information can be difficult for activities that include tens of sequences. This paper will discuss the design and usefulness of the tool, and also the overall activity review and approval process.
PlumeCAS is an all-in-one sample capture, volatile metering, and Capillary Absorption Spectrometer (CAS) system that can make sensitive isotopic and abundance measurements of H2O, CO2, CH4, and C2H6 in Enceladus’ plume to aid in the search for life. The instrument accumulates gaseous and icy phases during plume flythroughs in an aerogel and indium collector that preserves the integrity of small organics. The flight system would leverage a heritage hermetic sealing design to contain that sample, and a volatile metering system delivers aliquots to the CAS for analysis. There, hydrogen and carbon isotopic ratios of H2O and CO2 are measured to generate a baseline against which to compare isotopic ratios in CH4. PlumeCAS will also determine the abundances of CH4 and C2H6 to generate a CH4/C2H6 abundance ratio. These isotopic and abundance indicators are among the best-understood biosignatures available for measurement in Enceladus plume material. This analysis is nondestructive, and therefore the system has the potential to have additional gas analysis instruments added in series with the CAS. In this paper we describe the scientific motivation for the instrument, the breadboard hardware development and test campaign that validated this concept to TRL 4, and future work that may be done to better prepare the system for flight.
Stratospheric high-altitude balloons are non-extensible, sealed flexible structures designed to operate in the earth's stratosphere for extended periods. These balloons do not have propulsion engines, and their dynamics are entirely based on prevailing atmospheric wind conditions, making them vulnerable to being carried by wind currents. Station-keeping involves maintaining the balloon within a specific region for an extended duration. The standard control strategy utilizes atmospheric wind velocity variability with altitude, allowing a controller to predict the altitude with favourable wind velocities. In recent years, reinforcement learning-based station-keeping controllers have gained popularity. These controllers require extensive, realistic historical atmospheric training datasets to perform effectively. In the absence of such datasets, we propose a data-driven control strategy based on dual-mode extremum-seeking control (ESC) with vanishing oscillation for the navigation and station-keeping of high-altitude balloon platforms. Through simulation studies using real wind data from the National Oceanic and Atmospheric Administration (NOAA), we demonstrated that our proposed real-time optimization algorithm can successfully steer the balloon from one location to another without explicit knowledge of the prevailing wind dynamics.
The autonomous operation of an unmanned aerial vehicle (UAV) relies on reliable self-localization, which is typically achieved using global navigation satellite systems (GNSS). However, GNSS data can be unreliable due to effects of space weather phenomena or interference from GNSS jamming. To ensure accurate localization in such conditions, vision-based approaches for UAV positioning offer a potential alternative, though they often come with trade-offs in positioning accuracy or computational efficiency. In this paper, we present a real-time method for vision-based UAV self-localization that achieves GNSS-like accuracy. This approach involves extracting high-level semantic features from captured images and matching them to geo-referenced OpenStreetMap (OSM) data of the flight area. The global location of the UAV is then determined based on the matching results. We also compare and evaluate different metrics for measuring scene similarity to enhance the system's performance. Moreover, we demonstrate that even when OSM data is partially inaccurate, it can still be used to achieve accurate localization. This holds true even with a non-optimal neural network for segmentation and in environments with limited semantic features. The dataset used for evaluation will be made available with publication of this paper.
Space systems are complex, expensive, and challenging to design, develop, and launch, with limited options for servicing, upgrading, and extending their operational life. Decisions regarding their continued use or obsolescence are often subjective, influenced by budgetary or political factors rather than technical criteria. As systems age, performance may degrade, requirements evolve, and environmental factors shift, affecting system sufficiency. There is a need for a formal method to assess aging status and its relevance after space systems are deployed. The Technology Readiness Level (TRL), created in the 1970s, assesses technology maturity pre-deployment but is not designed to to monitor operational lifecycle stages, aging, or retirement. To fill this gap, the authors apply the extended Technology Readiness Level (eTRL), which builds on set theory to track system sufficiency through the intersection of requirements, performance, and environment. Changes in these factors can impact system relevance and lead to obsolescence. The application of eTRL is demonstrated through two case studies: the Hubble Space Telescope, which has exceeded its design life, and the retired NASA Space Shuttle. These examples highlight how the eTRL can effectively evaluate space systems' aging and contextual relevance throughout their operational lifecycles.
Advancements in deep learning and computer vision are enabling the development of expanded spacecraft capabilities. One field of interest is automated crater detection which has applications in terrain relative navigation, pose estimation, and planetary science. While continued development of learning-based crater detection algorithms (CDAs) has led to more accurate and performant models, there is currently a limited discussion on how these models might be integrated into future mission infrastructure. Specifically, we identify the deployment of CDA models onto resource-constrained, flight-like hardware and the interaction of CDAs with existing flight software as key areas of investigation. To this end, we first introduce a novel Lunar crater dataset based on digital elevation map (DEM) data and 1.2 million known crater positions, leveraging the Blender 3D software to render surface imagery with ground truth bounding box labels. We evaluate the You Only Look Once (YOLO) family of models on this dataset for crater recognition performance while providing runtime and memory analysis on representative flight hardware, consisting of a Teledyne radiation-tolerant LS1046-Space CPU and a Google Coral Edge TPU accelerator. We comment on the choice of activation function in the YOLO architecture as it relates to detection performance and inference time. Carefully considering model operations is essential because Edge TPU compatibility is paramount for near-real-time, onboard deep learning execution. Finally, we put forth an example implementation of YOLO CDA within the On-Board Artificial Intelligence Research (OnAIR) platform, a cognitive architecture for autonomous applications that can interact with flight software frameworks such as NASA's core Flight System (cFS).
This paper presents a three-step clearance process for flight control software. As a novelty, grid-based non-linear closed-loop simulation testing has been added within a unit test framework as a validation step. This new feature allows developers to assess compliance with requirements quickly on a very detailed level. Within this work, the extensive testing and the validation implemented for this unit testing framework are presented. Based on high-fidelity closed-loop simulation scenarios, the unit tests, including actuator and sensor dynamics, cover all nominal flight, disturbance, and fault conditions. Furthermore, test coverage is addressed and investigated as an indicator of the test framework quality. In the second and third steps, optimization-based clearance methods and Monte-Carlo simulation techniques assess the performance of the flight control laws in the presence of a worst-case aerodynamic parameter variation. The Monte Carlo simulation validates the worst-case scenario over the entire flight envelope and for all gust disturbance scenarios considered. The presented application example is the German Aerospace Center's High Altitude Long Endurance (HALE) platform HAP-ALPHA, which is used to illustrate all steps of the presented validation and verification (V &V) framework. The design and development of the flight control software development for the HALE aircraft has been discussed in previous work and has the novelty of considering the aircraft's flexible modes for the primary flight control design. The V &V framework assures compliance with functional, performance, and software quality requirements, allowing entry into an aircraft's flight testing phase.
This report tests an extension of Gooding's Method for Initial Orbit Determination (IOD) to process N line of sight (LOS) measurements for space-based observations of a target taken in a near co -planar orbit relative to the observer. Additional error frames are added for measurements 2 through N-1 to extend the gradient used in a Newton-Raphson iterative refinement formula from 2 target functions to 2*(N-2). The NGooding algorithm utilizes a least-squares approach to perform the iterative refinement on a non-square matrix. Measurement errors and process noise are not considered in this study, and dynamics are represented from simple Keplerian orbits. Line of sight observations are processed as unit vectors and not converted to error angles as seen in other N-Gooding approaches. A sample of 100 orbits were tested for N = 3 through N = 24 observations. The accuracy of the N-Gooding algorithm significantly improves on the original Gooding's algorithm in consistently fewer steps. The resulting N-Gooding estimate shows an average of 80% improvement in position error as compared to the traditional Gooding algorithm, with average errors of 0.0467 km and 0.2378 km respectively. Both accuracy and precision improved with more observations for a majority of target orbits, with standard deviations reduced by more than 60%. The number of steps to converge to within an error frame tolerance of +1- 1E-8 km was reduced by an average of 99% The resulting convergence plots show that the estimate improves according to a logarithmic trend. Future work includes testing the effects of process and measurement noise on the accuracy and convergence of the N-Gooding algorithm. Classification of near co -planar orbits that yield optimal and predictable results against target that are less likely to converge will also be studied. Last, comparison of N-Gooding to other IOD procedures such as those based on Physics Informed Neural Networks is of particular interest.
This paper presents flight results and lessons learned from optical angles-only navigation of a satellite swarm, conducted during the Starling Formation-Flying Optical Experiment (StarFOX). StarFOX is a core payload of the NASA Starling mission, which consists of four propulsive CubeSats launched in July 2023. Prior angles-only flight demonstrations have been limited to a single observer and single target, and have relied upon a-priori target orbit knowledge for initialization, translational maneuvers to resolve target range, and external absolute orbit updates to maintain convergence. StarFOX removes these limitations by applying a new angles-only architecture called the Absolute and Relative Trajectory Measurement System (ARTMS), which integrates novel algorithms for image processing, batch orbit determination and sequential orbit determination. During StarFOX experiments from December 2023 to May 2024, images from on-board star trackers were used to perform multi-target and multi-observer relative navigation; to autonomously track and initialize navigation for unknown targets; and to perform simultaneous absolute and relative orbit determination without GPS. Relative positioning uncertainties (1s) of 1.3% of target range are achieved for a single observer, reduced to 0.6% with multiple observers (without orbit control maneuvers). However, in-flight conditions proved more challenging than anticipated with respect to swarm visibility, image signal-to-noise ratios, image time-tag synchronization, and overall measurement reliability. The impact of these conditions on performance is studied by comparing in-flight telemetry to telemetry produced by an ARTMS digital twin running on the ground, using both in-flight measurement data and synthetic measurement data from pre-launch simulations. Usage of the digital twin facilitated efficient troubleshooting and mitigation via in-flight software updates. Upcoming StarFOX+ experiments in 2025 will extend ARTMS with additional capabilities, such as opportunistic detection and identification of passing resident space objects, which are preliminarily demonstrated via proof-of-concept simulations using StarFOX flight data.
INvestigation of Convective UpdraftS (INCUS) is a NASA Earth Venture Mission (EV-M) managed by JPL for the Colorado State University (CSU) PI. The INCUS mission will be comprised of a constellation of three observatories, each consisting of the spacecraft bus, and the Dynamic Atmospheric Radar (DAR). One observatory will also host the Dynamic Microwave Radiometer (DMR), based on the TEMPEST-D technology demonstrator. DAR is an evolution of RainCube (Radar In a Cubesat), with 7-beam cross-track scanning capability. The three observatories will fly in low-Earth-orbit (LEO) in a train formation to make measurements of tropical convective storms at three different times (0, 30 and 120 seconds). The temporal sampling of the DAR reflectivity profiles is used to measure reflectivity changes with time, allowing the detection of convective updrafts and the estimation of their transport of air and water between the surface and upper troposphere, known as Convective Mass Flux (CMF), and elucidates how this transport varies with storm type and throughout the storms' lifecycles. The INCUS measurements will increase our understanding of how, when, where, and why storms form, and will allow us to better model and predict the role of convective storms in current and future climates. This paper focuses on how the lessons learned from on-orbit flight operations as well as the design heritage and implementation and test experience of RainCube technology demonstration were leveraged in developing the concept, system requirements and implementation approach for the INCUS mission's DAR instrument. We also present the programmatic considerations associated with translating a technology demonstration mission to a higher class (class C or D) mission and the trade-offs considered during such development. RainCube was a 35.75 GHz Ka-Band radar with 0.5 m deployable antenna in a 17 kg 6U CubeSat form factor. RainCube was successfully launched to ISS on May 21, 2018, deployed to an orbit of 350 km in August, 20218 and completed a 2-year extended mission in LEO on Dec 24, 2020. INCUS is a 35.75 GHz Ka-Band radar with a 1.6 m deployable antenna in a 130 kg Small Satellite and has a target orbit of 575 km. INCUS is scheduled to launch in 2026.
Autonomous in-flight inspections of aircraft battle damage by unmanned aircraft systems (UAS) can provide critical information to operators while improving safety and eliminating the need for visual in-flight inspections with other manned aircraft. This work centers on the use of an UAS with commercial off-the-shelf (COTS) infrared (IR)-based sensors to autonomously identify battle damage on aircraft. The proposed solution employs a convolutional neural network (CNN), Simultaneous Localization and Mapping (SLAM), and a pixel-to-world point converter to detect, classify, and localize aircraft damage in real-time and onboard. This paper presents an experiment that autonomously detected and identified aircraft battle damage on a T -37 airplane wing in day and night conditions with a 58.02% success rate on average. Accurate localization, though verified experimentally with different sensors, was not fully tested due to mediocre estimations of sensor intrinsics. This methodology has promise in many applications related to battle awareness due to its improvements over standard red, green, blue, depth (RGB-D) sensor packages under poor visibility conditions. Potential future applications of this work include remotely piloted aircraft (RPA) using existing sensor packages for aircraft monitoring and fighter aircraft wingmen passively inspecting one another. This technology could also be used by commercial aircraft and spacecraft to quickly and confidently identify structural damages or other anomalies. Future work includes increasing the quality and quantity of training for the CNN, further characterizing the sensors, adapting the system to inspect moving aircraft, and fusing the IR sensors with standard color cameras.
The LunaNet Interoperability Specification (LNIS) has been created and released to the public, enabling international commercial and government lunar systems to have a common baseline to work together in forming complex lunar mission networks. The LNIS specifically defines a set of data service protocols that offer multiple options for different types of network users and providers to work together. This includes real-time frame delivery services similar to those classically used by missions, as well as both real-time and store-and-forward networking that will be important as the number of mission users and complexity of lunar mission operations increases. The NASA Lunar Communications Relay and Navigation Systems (LCRNS) project requires lunar relay satellite services to implement onboard processing and networking capabilities in some ways similar to LEO mega-constellations that offer Internet access, but also going beyond those capabilities to meet unique lunar mission needs (e.g. store-and-forward services, LNIS messaging, etc.). Onboard processing will be necessary for a number of different protocols that are included in the LNIS, including support for data transfer over lunar proximity radio links based on CCSDS framing and CCSDS Encapsulation Packets, networking via IP and DTN Bundle Protocol, a suite of DTN Convergence Layer Adapters, and LunaNet messaging services. This paper provides an overview of our work in contributing to and defining the LNIS service interfaces and LCRNS requirements related to onboard processing for data services. This includes exploration of the key aspects of lunar networking concept of operations, technology trade studies, protocols for trunking, applicability and scope of different protocols, messaging, security considerations, and realistic implementation and deployment constraints. While the technology for new lunar relay networks can build upon the recent advances in LEO constellations, there are unique new challenges related to network management, onboard processing hardware/software system constraints, security, and quality of service for human spaceflight missions. As multiple international systems are planned to be put online in the coming years based on the LNIS and derived data services, this work will have a lasting impact that supports the evolution from single-vehicle user missions into scenarios where many surface and orbital assets are able to collaborate over multi-hop networks operated by diverse providers.
Autonomous in-space assembly systems support deep space exploration and sustained presence in space by enabling the development of large-scale infrastructure. Assembled mechanical metamaterials are a promising technology for space applications due to their low mass density with high mechanical performance. Previous building block based material and structure designs include monolithic units and numerous styles of composition. Component geometry such as strut-and-node designs have been investigated for their high packing efficiency in launch vehicles. Deployable building blocks provide a new paradigm for increased packing efficiency during transport. In this paper, we explore a novel deployable building block design using coiled deployable booms. Coiled deployable boom technology is used extensively in space structures for their lightweight properties and low power requirements. They are compact in the stowed configuration and have a high stiffness to mass ratio when deployed. We present the design and performance characterization of a deployable, adaptive, structural building block for assembled metamaterial systems. The building block is designed with tape springs that serve as the strut members between attachment nodes. This building block design has a high packing efficiency for transport, tunable strut design for custom interfaces and mechanisms, and allows for efficient repair of areas deep inside a structural volume, providing a highly adaptive system for assembly, maintenance, and reconfiguration. The structural performance of the building block is analytically estimated then evaluated with physical experiments, and two folding configurations of the building block prototype are demonstrated. A discussion of lessons learned and future design recommendations is also presented. This system of adaptive structural modules for robotic assembly aims to enable complex assembled structures and space infrastructure.
A review of operations and performance for the electric propulsion (EP) system on the Psyche spacecraft during the first year of the mission is presented. After a successful initial checkout phase, the EP system has been in near continuous operation since late March 2024, the beginning of a roughly six year long Cruise phase. Usage statistics to date on all four thrusters and both power processing units (PPUs) are presented. Notable trends in electrical and thermal performance of the thrusters, feed system, and PPUs are discussed along with unexpected observations and risk mitigation activities. The propellant gauging method is reviewed and a to-date trending of xenon mass with uncertainty is presented. Improvements to the spacecraft momentum management strategy made since launch are detailed alongside measured total torque on the spacecraft during thrusting. The new thruster switching plan for the rest of the six year Cruise and an operations-centric propellant budgeting strategy are presented for the first time here.
NASA's Artemis program and Moon to Mars objectives target development of a sustained human presence on the Moon. This drives the need for understanding what technologies and procedures are critical for long-duration surface habitation, in both a lunar and Martian environment. Earth-based analogs provide the ability to test mission components in comparable environments. However, there are a limited number of analog facilities, each with high-fidelity approximations of different features of the target environment. The limited availability of analog spots and associated high costs of testing make efficient use of analogs as a research platform critical to provide useful data for space exploration. Here, we propose a framework to evaluate the merit and need for testing a given experiment in different analog facilities. Multiple criteria decision analysis techniques utilizing analytic hierarchy process calculations are used to evaluate the compatibility of each analog facility's level of feature approximation with a given experiment or mission profile. 24 simulation features such as terrain, isolation, mission control, and available technology are used to evaluate facility approximation and analog needs of researchers. The experimental value, or merit, of a given experiment is quantified according to NASA's targeted knowledge gaps as ranked in the Science Technology Mission Directorate's shortfall list and the Human Research Program risk assessment, and the proposed technology/human readiness level increase resulting from the experiment. A prospective utility assessment using this method is conducted to evaluate optimal facilities for the analog testing of two representative experiments: a lunar ISRU energy production technology, and a psychosocial crew teaming mission plan. This selection methodology framework quantifies the merit of conducting a given experiment in various analog environments. It enables analog selection by researchers to target the best possible facility for their work, and enhances the ability of facilitators to select experiments that utilize the unique capabilities of their analog. The efficient use of analog resources via optimized experiment and facility selection will enable improved and rapid advancements to the technologies deemed most critical to test prior to in-situ integration.
There is a growing need to make software and hardware development efforts run in parallel to reduce cost and time. Creating a digital twin of a spacecraft's peripheral interfaces allows for earlier flight software testing, without a need for physical hardware. This paper explores the creation of a digital twin for the Cyclone Global Navigation Satellite System (CYGNSS) spacecraft using Renode and Verilator toolsets. In addition to the cost and time saving benefits, a digital twin also enables initial hardware/software integration testing. Bugs that may normally be found at the beginning of hardware integration can be squashed earlier before the mission hardware asset is available. Additionally, common issues in software such as clocking domains, reset architecture, and sequencing issues that can cause damage to the hardware can be discovered without risk. Several obstacles prohibited fast development of the virtualized hardware. One of which was, working with a model of a system as complex as CYGNSS led to issues related to timing. Running simulations took anywhere from 1/10th real time speed to 1/1000th. However, simulation time was greatly reduced to near real-time by incorporating stub-like modules to mimic Field Programmable Array (FPGA) functionality. The resulting approach found would be useful in improving flight software fidelity during development of the system, but would not be ideal for simulation of existing, complex systems.
As Proliferated Low Earth Orbit (pLEO) satellite constellations increase in popularity for providing low latency global communications networking, their vulnerability to dis-ruptive threats to deny use is also growing. Existing routing schemes for pLEO are either centralized terrestrial schemes that typically rely on knowledge of link and node states in the entire constellation, or distributed low-complexity space-based schemes that rely on factors such as orbit geometry or geographic location of endpoints. In this paper, we propose a hybrid routing scheme called Primary Routing Aware Backup Routing (PRABR). PRABR is a framework to incorporate a centrally planned primary routing algorithm alongside a distributed backup routing algorithm and intelligently switch between them based on local observations of a router. We use the ns-3 based Hypatia simulator to show that PRABR significantly outperforms primary routing alone under satellite failures while maintaining the latency minimizing properties over the backup routing algorithm alone. We also present several example routes and discuss the effects of satellite failures and PRABR on such routes.
In this work, we develop a solution to the initial orbit determination (IOD) problem from sequential inter-star angle measurements. With the constantly increasing interest in deep space missions, alternative methods of navigation are gaining more and more attention. Scenarios where classic Earth-based navigation sources are not available require new spacecraft to be able to autonomously detect their own orbit using information accessible in unfamiliar environments. Throughout the Solar System, stars are one of the few sources of information which is always available and the effect of stellar aberration may be used to infer the velocity vector of the orbiting body. In the problem analyzed in this work, the spacecraft measures a single inter-star angle at each time, with the star pair potentially changing from observation to observation. The solution is obtained in two steps: the first aims at providing an initial guess of the orbit by modeling the velocity vector with the Lagrange interpolating polynomial. The second step refines the initial guess by accounting for the two-body dynamics through a propagation approach based on Euler's finite rotation formula. This propagation method reveals to be especially convenient in the StarNAV context, where the knowledge of the position vector is superfluous and its propagation can be avoided.
NASA's Europa Clipper spacecraft will carry nine science instruments to investigate whether Jupiter's moon Europa has an environment that could support life. Europa Clipper is the largest spacecraft NASA has flown for a planetary mission. With solar arrays fully extended the spacecraft spans over 30 meters. At launch, nearly half of Europa Clipper's mass is propellant which will be consumed during travel to Jupiter, inserting the spacecraft into orbit around Jupiter, and to perform almost 50 flybys of Europa. Europa Clipper's propulsion system has 24 bipropellant engines which will process nearly 2750 kg of fuel and oxidizer during the course of the mission. During propulsion system operation liquid droplets of unburned fuel and oxidizer are expelled from the engines. An empirical bipropellant plume contamination model that was originally developed for International Space Station thrusters was applied to investigate droplet effects on the Europa Clipper spacecraft. Unburned propellant droplets can cause damage to spacecraft materials by chemical reactions and mechanical damage exhibited by pitting surfaces. Bounding estimates for propellant droplet fluxes to Europa Clipper surfaces were obtained using the empirical model. Based on material testing results, estimates of damage to materials used for thermal blankets, solar arrays, and instruments were assessed to determine the impact to spacecraft thermal, power and science systems. Where needed, design changes were made to mitigate deleterious effects caused by propulsion system operation.
Using Vitis AI, a Xilinx framework for executing machine learning models on heterogeneous computing hardware, this work benchmarked the capability of a Xilinx Versal AI Core system-on-a-chip (SoC) to execute a runway detection segmentation neural network. This demonstration was performed to support an NRL developed algorithm for performing landmark based orbit-determination. The radiation tolerant Xilinx Versal Core AI XQR provides an option for high-throughput on-orbit processing to support applications such as neural networks for electro-optical sensor processing. A Xilinx VCK190 evaluation kit was used to perform the benchmarking. The model used in this effort was a modified DeeplabV3+ segmentation model that performed well using a limited data set of 160 training images and 40 test images. The DeeplabV3+ model was altered to allow for a Xilinx Deep Processing Unit (DPU) instantiation of the model that did not rely on custom implementation of any layers in the network. This was done to increase throughput and ease the implementation of the model on the Versal. The data set was built from imagery captured by the WorldView-2, WorldView-3, and GeoEye-1 satellites. In evaluation of the accuracy, the mean-intersection-over-union (mIoU) was determined to be roughly 0.6772 on the floating-point model following training. The model was then quantized from floating point to INT8 in order to allow for FPGA compatibility, and then compiled for execution using Vitis AI on the Xilinx Versal VC1902 SoC hosted on the VCK190 development board. The resulting model was run using the Vitis-AI API via a python script, and the resulting mIoU was found to be 0.6757. The maximum throughput achieved in this configuration was 70.108 FPS using 7 threads for evaluation on the VCK190.
