This paper describes the design and development of the Multi-Angle Imager for Aerosols (MAIA) instrument from concept through delivery. Selected in 2016 in response to the third NASA Earth Venture Instrument (EVI-3) opportunity, MAIA enables the investigation of links between different types of airborne particulate matter (PM) and adverse human health impacts such as cardiovascular disease, respiratory disease, pre-term delivery, and low birth weight. The instrument optical system features a push-broom UV/VNIR/SWIR spectropolarimetric camera with a passively cooled focal plane module and a pair of novel photo-elastic modulators to measure the radiance and polarization of sunlight scattered by atmospheric aerosols, from which the abundance and characteristics of ground-level PM are derived. The camera is mounted on a two-axis gimbal that allows multi-angle pointing, frequent target revisits, and in-flight calibration capabilities. The EVI approach presented unique challenges to instrument formulation such as designing for spacecraft “hostability,” using standardized requirements in the absence of defined spacecraft interfaces and environments, and adopting simple approaches to operability and fault protection. Implementation challenges included managing schedule and logistics during the SARS-CoV-2 pandemic, late changes to the gimbal design, completing an extensive camera calibration campaign, and expediting hardware and software modifications once the host spacecraft was selected. Several key trades and a descope undertaken during the design process will also be described. An overview is provided of instrument integration and test, which was completed in 2022. In 2023, NASA and the Italian Space Agency (Agenzia Spaziale Italiana or ASI) agreed to implement the MAIA mission as a joint NASA-ASI partnership with ASI contributing the spacecraft as well as the launch vehicle. The MAIA mission is currently planned to launch on a Vega-C launch vehicle from the Guiana Space Center in 2027.
The NASA Psyche Discovery Mission is currently cruising towards a rendezvous with the asteroid (16) Psyche, the largest M-class asteroid in the main asteroid belt [1]. The spacecraft instrument suite consists of a magnetometer, gamma ray and neutron spectrometers, multispectral imagers, and radio science experiments, all designed to unravel the history of (16) Psyche [2,3]. The resulting data products generated by the mission (e.g., images; spectroscopic, magnetic, and gravity field data; shape models; geologic maps, etc.) will be delivered to the Small Bodies Node of the NASA Planetary Data System (PDS) for long-term archiving [4]. The Psyche mission’s Science Data Center (SDC), part of the JPL Psyche Mission System, is the point of contact for all data sharing and archiving activities. Here we describe the design and implementation of the SDC, which was guided by many factors: supporting mission requirements related to data warehousing; performing data dissemination and archiving; adhering to federal, NASA, and ASU cybersecurity guidelines; and following industry best practices. Given the long baseline of the mission (launch in October 2023 and arrival at the asteroid in mid-2029), the system needs to be easily maintainable and upgradable during the mission’s lifetime. The SDC is located on the Tempe campus of ASU, with connections to the mission’s Ground Data System at JPL, and is available to the Psyche team via a web portal utilizing purpose-built tools. The SDC leverages heritage tools, concepts, and lessons learned from previous ASU instrument operations, such as for the Lunar Reconnaissance Orbiter Cameras on LRO, the Mastcam cameras on the Mars Science Laboratory rover, and the Mastcam-Z cameras on the Mars 2020 rover. We describe the heritage, principles, and requirements that guided the design and initial development of the Psyche Science Data Center, including real-world examples of the SDCs data portal, RESTful interface, in-house scripts, and early data products. The design of the SDC is centered around a relational database, with a schema to model the many files that are ingested and tracked, as well as their relationships to the PDS bundles being aggregated and delivered. The SDC disseminates data products to the Psyche Team for science investigations through a web-based portal, which also includes a RESTful interface that allows team members to upload, download and search data using in-house scripts. The three instrument teams make heavy use of the RESTful interface for uploading their PDS products. The web portal also makes mosaics and individual image products available to the team using Web Mapping Service technology. Most of the tools developed for the SDC are written in Python, using virtual environments to minimize the need to configure and maintain Python at the system level. These adhere to the UNIX philosophy of software development: make each program do one thing well, expect output (when generated) to become input to another tool, and test early and refactor as needed. References: [1] Elkins-Tanton+, Space Sci. Rev., 218, 2022. [2] Dibb+, AGU Advances, 5, doi:10.1029/2023AV001077, 2024. [3] Polanskey +, Space Sci. Rev., submitted, 2025. [4] https://pds-smallbodies.astro.umd.edu
This paper reports on our mathematical framework of an autonomous, vision-based interception algorithm for non-cooperative targets, its implementation on three separate types of mobility platforms, and experimental results from field testing. The three types of mobility platforms include an unmanned aerial vehicle, a four-wheeled skid-steered ground vehicle, and a 700-pound spacecraft simulator that moves on a flat floor on air bearings using 8 thrusters. We use only a monocular camera stream with fiducials for target tracking. The same approach is validated on the three types of platforms to understand generalizability by addressing platform specific idiosyncrasies and dynamics. Our algorithmic approach is based on (1) a nonlinear Extended Kalman Filter to estimate the relative target pose amid intermittent measurements, (2) an uncertainty-aware motion predictor that propagates a plausible target trajectory conditioned on its history, and (3) a receding-horizon planner that solves a constrained sequential convex program in real-time to guarantee a kinematically feasible and time-efficient interception path. It makes minimal assumptions about the environment, operates entirely in the local observer frame without global information, and relies on a single sensing modality—a monocular camera stream. The operating regime also assumes limited observability due to partial fields of view, sensor dropouts, and target occlusions. For target detection we deploy a classical approach for nominal cases and a learning-based approach to address off-nominal image quality. We incorporate platform-specific details such as degrees of freedom and vehicle dynamics through modular interfaces. Ground rover experiments involve (1) two rovers rendezvousing and stopping at a prescribed distance and orientation and (2) leader-follower scenarios with a passive follower rover maintaining a prescribed relative pose offset from its leader. The unmanned aerial vehicle experiment demonstrates a single UAV identifying a dynamic target fiducial and landing on it. The experiments with the spacecraft simulators involve two such bodies using thrusters for dynamic rendezvous and station keeping. We evaluate and compare interception pose errors, success rates under different target motion profiles, and computational latency on multiple embedded processors (an Nvidia Jetson Orin, a ModalAI VOXL2, and a Raspberry Pi 5). In each scenario the target’s motion is dynamic and non-cooperative exhibiting stochastic behavior. Results show excellent performance, for example, with the ground vehicles demonstrating 3-5 centimeter errors over tens of meters of relative travel on natural terrain. The paper will also discuss the impacts of environmental factors and steps taken to mitigate these issues.
The proliferation of space debris has become a matter of grave concern for new and existing orbiting assets, leading to potential collisions and losses. One approach for mitigating the impacts of space debris involves the use of satellites with robotic manipulation arms mounted on them. The approach for capturing debris with such a system involves going to a rendezvous orbit with a previously identified debris object, using a vision-based grasp planner to find an approach trajectory for the robotic arm, and finally controlling the attitude of the satellite along with the joints of the robotic arm and gripper to execute a successful grasp. Building upon our previous work on grasp planning for satellite-mounted manipulators, after the orbit rendezvous stage, we propose a whole-body control strategy using reinforcement learning (RL) to control the combined satellite and robotic arm system. In order to execute debris capture in a closed-loop manner, we employ a lightweight vision-based grasp planner that allows for real-time visual servoing, as well as tactile sensing on the robotic arm’s end-effector to prevent the target object from slipping. The reinforcement learning model, once trained, would be able to operate with lower computation costs than classical approaches such as MPC. We demonstrate the efficacy of this closed-loop RL-based controller for debris capture in a high-fidelity physics simulation environment and compare its performance with classical controllers.
A radio was designed to support simultaneous co-channel jamming and communications using an in-band Simultaneous Transmit and Receive (STAR) antenna array followed by two layers of adaptive estimator/subtractors. These cancellers use a reference signal obtained by tapping off a small fraction of the transmitted jamming signal. The estimator/subtractors can suppress other locally transmitted signals if they are tethered to the receiver to provide reference signals. The output of the implemented local interference cancelling front-end resembles a four-channel array, and array interference-nulling techniques can be applied to the four-channel output to suppress untethered-interference signals. Due to the preceding adaptive cancellers, the untethered-interference nulling algorithm used cannot require array calibration. The received communication signal has a known format and contains known symbol sub-sequences that can be used as training data. The proposed algorithm for untethered interference cancellation with signal training data and without array calibration maximizes the output signal-to-interference-plus-noise ratio using the training sequence to estimate Minimum Variance Distortionless Responce (MVDR) type array weights without knowing the signal Angle-of-Arrival (AoA). Here we call it MVDR for Uncalibrated Arrays (MUA). This paper describes the receiver design and presents simulation results for the MUA algorithm applied to the four-channel output. It is shown that with ideal self-interference cancelling untethered-interference nulling performance is good whenever the input Interference-to-Noise-Ratio is high enough for the algorithm to implicitly estimate the array interference response, and not so high that the algorithm cannot reliably detect signal training data, provided the interference and signal AoAs are at least 45° apart. Signal to Interference Ratio improvements as high as 50 dB have been observed when untethered interference is strong.
Over the last few decades, many planetary missions have relied on Radio Science signals and observations to probe the planets and improve our understanding of the solar system. Recently, JPL has utilized NASA's Gravity Recovery and Interior Laboratory (GRAIL) radio science as Signals of Opportunity (SoOP) to map the moon's ionosphere and investigate the measurements for the future Lunar Radio Occultation (RO) mission, which aims to study lunar dust and surrounding plasma interactions. This research will provide an assessment of the GRAIL radio science signal noise patterns corresponding to different locations in Geocentric Solar Ecliptic (GSE) coordinate, solar plasmas conditions, and orbit dynamics of radio occultation experiments for the new scientific observations derived using GRAIL RO data; the statistical analysis of the noise patterns, estimated using the actual GRAIL RO data, will be in comparison with the uncertainties derived using analytical models. The research findings provide a reference for the scientific community to design future lunar Radio Occultation Missions, improving the understanding of constraints and errors in these new scientific measurements made by GRAIL-like Radio Science instruments.
Chascii is currently developing the inter-spacecraft omnidirectional optical communicator (ISOC) to provide fast connectivity and navigation information to small spacecraft forming a swarm or a constellation in LEO and cislunar space. The ISOC operates at 1550 nm and employs a dodecahedron body that holds six optical telescopes and 20 external arrays of detectors for determining the angle of arrival of the incoming beams. Additionally, the ISOC features six fast avalanche photodetector receivers for high-speed data rate connectivity. The ISOC should be suitable for distances ranging from a few kilometers to a few thousand kilometers. It will provide ubiquitous coverage and gigabit connectivity among smallsats, forming a constellation around the moon. It will also offer continuous positional information among these spacecraft, including bearing, elevation, and range. We also expect the ISOC to provide fast, low-latency connectivity to assets on the lunar surface, such as landers, rovers, instruments, and astronauts. Chascii is currently developing a lunar ISOC, including all its transceivers, optics, and processing units. We are developing the ISOC as a key candidate to enable LunaNet. In this paper, we will present the development status of the ISOC as well as link budget calculations for operations around the moon using pulse-position modulation. We will present experimental results of angle-of-arrival testing using various experimental apparatus. We will also present connectivity results obtained with two ISOCs, including measured bit-error tests under different conditions. We will discuss our ISOC development roadmap, which includes LEO, GEO, and lunar missions, spanning the 2026-2029 timeframe. We believe the ISOC, once fully developed, will provide commercial, high-data-rate connectivity to future scientific, military, and commercial missions around LEO, cislunar space, and beyond.
Signals integration is used in digital communication systems with data fusion to enhance the performance and reduce the multipath effect. The two main approaches for signals integration in digital communication systems with data fusion are full and semi-full signals integration. In full signals integration systems, there are multiple receivers producing very large number of bits and the entire signals integration system is closely resembled analog multiple receiver implementations. This approach achieves the optimum performance at the expense of high cost and complexity. In semi-full signals integration systems, only few numbers of bits are used after preliminary processing of signals at each individual receiver. This method could reduce system complexity and cost at the expense of overall performance degradation. This paper provides performance analysis of full and semi-full signals integration approaches in digital communication systems in case of non-coherent frequency shift keying (NCFSK) receivers with Gaussian noise and Rician fading stochastic model. The performance loss due to semi-full signals integration is analyzed for different number of information bits.
Satellite constellations and swarms are a concept that is rapidly becoming more common in the world of space systems. Organizations ranging from the government to private internet providers are starting to utilize multiple satellites working together to accomplish their mission, often with the ability to communicate with one another in orbit. This survey aims to cover and provide a brief summarization and overview of the current world of satellite systems and constellations and the current state of their security using open source information. Focus is put on systems currently in development or deployed, real world cyber attacks against satellites, research being conducted on satellite testbeds for cybersecurity, and research that is being conducted for theoretical attacks against satellites. Systems covered include Starlink, Kuiper, ViaSat, GNSS, those from NRO, NASA, SDA, and several more. Each system is briefly compared and their security posture is analyzed from the perspective of an attacker and defender. It is expressed in this survey that several of these systems offer robust security, but several may contain gaps that attackers can exploit. Additionally, some of the experimental systems may offer new attack vectors, and thought is given to how those could impact the security of the constellations. The goal of this survey is to highlight the significance of cybersecurity in satellites, and where attackers might direct their attention in the future. It also briefly provides advice and resources for cybersecurity researchers to enter the field, with hopes of providing avenues for researchers to get started in a very challenging field.
This paper details the design, verification, and validation (V&V) of the Hardware Electronic Real-time Message Exchange System (HERMES), a satellite surrogate communication testbed (SSCT). HERMES leverages commercial off-the-shelf (COTS) hardware to emulate a deployed satellite communication subsystem. The system comprises a small and low-cost ground-station (GS) setup testbed for evaluating the data-link communication between the satellite and GS. HERMES weighs the capable open-source ground-station user interface OpenC3 COSMOS for sending packets while offering a firmware for emulating the responses from the satellite flight computer. The results show that HERMES offers excellent value to student-led CubeSat programs. One advantage of HERMES is the possibility of testing the communication between the satellite and GS without waiting for a functioning engineering model of the satellite and the onboard computer (OBC) subsystem firmware. This is achieved because the SSCT testbed offers an open-source surrogate satellite firmware deployed to the COTS microprocessors that emulate the satellite's OBC behavior. Another advantage of HERMES lies in its standardization of communications, which uses standard packetization. This standard lays the foundation for the early adoption of mission packet definitions adaptable to various mission objectives and payloads. Therefore, while HERMES can receive and decode uplink commands sent from COSMOS/GS and handle the packets to produce and send downlink telemetry back to the COSMOS/GS, it enables the GS development to advance in parallel to the satellite hardware and flight software. The radio frequency communication is transmitted, between two SDRs, while an embedded microprocessor emulates the flight radio leveraging GNU radio software. The experiment results demonstrate all packets sent to and from HERMES are being received and correctly interpreted for all configurations. The latency of the packet communication process between COSMOS and the GS hardware was calculated and saved as part of the meta-data for each experiment. Furthermore, HERMES was evaluated using a flight unit from the upcoming Virginia Tech's CubeSat UTProSat-1's OBC, slated to launch in 2025. The overall comparison results demonstrate the potential of HERMES as a surrogate COTS satellite subsystem and ground-station testbed for communication V&V with OpenC3 COSMOS.
A tracking system can efficiently track a target if the motion model used by the system match the target’s motion model. The Interacting Multiple Model (IMM) algorithm is commonly employed in tracking systems for this purpose. IMM uses a set of models to represent the possible evolution of the target’s state. Since the target’s motion can switch between different modes, the IMM algorithm must switch accordingly. The switching between models is governed by the Transition Probability Matrix (TPM), which plays a key role in determining both the estimation accuracy and the response time of the tracker. In the conventional Interacting Multiple Model (CIMM) algorithm, the TPM is predefined and set heuristically. The diagonal elements of the TPM represent the probability that a model will continue in its current state, which directly affects the accuracy of the estimate. A higher value for the diagonal elements typically leads to more accurate estimates in the case of model matched filtering. However, larger diagonal elements also result in slower model switching, as the off-diagonal elements determine the speed at which models switch during changes in the target’s motion. This creates a trade- off between accuracy and responsiveness, limiting the performance of the CIMM algorithm. To address this limitation, this paper proposes a likelihood-based approach (ATPM L) for updating the TPM. In the proposed method, the likelihood of each model is computed with respect to the current scan measurement and the TPM update factor is derived based on these likelihoods. This likelihood-based approach satisfies both accuracy and switching speed requirements of TPM in IMM algorithm. This algorithm also offers a faster model switching mechanism compared to all existing TPM update algorithms.
Flexible spacecraft structures present significant challenges for physical and control system design due to nonlinear dynamics, mission constraints, environmental variables, and changing operational conditions. This paper presents a data-driven framework for constructing reduced-order surrogate models of flexible spacecraft using the method of Dynamic Mode Decomposition (DMD), followed by optimal sensor/actuator pair placement. High-fidelity simulation data from a nonlinear flexible spacecraft model, including coupled rigid-body and elastic modes, are captured by defining a mesh of nodes over the spacecraft body. The data-driven methods are then used to construct a modal model from the time histories of these node points. Optimal sensor/actuator placement for controllability and observability is performed via a nonlinear programming technique that maximizes the singular values of the Hankel matrix. Finally, the sensor placement and dynamics modeling approach is iterated to account for changes in the dynamic system introduced by sensor/actuator physical mass. The proposed methodology enables initialization of physical modeling without requiring a direct analytical model and provides a practical solution for onboard implementation in model-based control and estimation systems. Results demonstrate optimal design methodology with substantial model-order reduction while preserving dynamic fidelity, and provide insight into effective sensor-actuator configurations for estimation and control.
Objectives General Matrix-Matrix Multiplication (GEMM) and General Matrix-Vector Multiplication (GEVM) operations are critical in multifunction space communications applications, ranging from free-space optical integrated sensing and communications (ISAC) to ISAC-enabled satellite systems. The underlying components of such systems include beamforming, channel estimation, interfer- ence mitigation, and adaptive filtering, among others. This paper aims to develop an efficient implementation of GEMM and GEVM using a Multi-Instruction Multi-Datapath (MIMD)-based Domain Adaptive Processor (DAP). The goal is to achieve high throughput, low latency, and energy-efficient computation while maintaining scalability and reconfigurability. Methods The DAP integrates a runtime-reconfigurable systolic ar- ray within the DASH SoC platform. Unlike SIMD processors, the MIMD architecture enables fine-grained control, parallel execution, and tailored dataflow. A row-stationary, column-streaming strategy is introduced to reduce memory overhead and support pipelined computation across processing elements (PEs). A case study ex- amines GEMM for radar-communication interference mitigation, focusing on a 4×64 by 64×4 cross-covariance matrix mapped to an 8×8 PE array. Analytical modeling estimates latency, compute time, throughput, power, and energy efficiency across varying array sizes, geometries, and matrix dimensions. Execution time is broken down into instruction loading, data loading, and computation/routing costs. Results Scaling PEs improves latency up to 3× (with diminishing returns beyond 16 PEs). Computational cost grows nearly linearly, showing ∼2× improvement at 16 PEs. Throughput saturates at 8 PEs but increases steadily (≈2×) with more columns or PEs. Power consumption remains stable (≈1.08–1.16× growth under higher workloads), highlighting efficient streaming dataflow. Geometry affects performance: with 8 PEs, latency gains range from 1.25× (4×2) to 4× (1×8); with 16 PEs, gains range from 1.16× (4×4) to 2.11× (2×8). Diagonal mappings introduce higher overhead due to extensive PE usage. Scalability analysis shows consistent 1.9–2× gains in latency and throughput when input dimensions double, while energy scales linearly with workload. Conclusion The proposed MIMD-based DAP achieves significant improvements in GEMM/GEVM performance, delivering up to 8048.7 GOPS/W, surpassing contemporary accelerators such as Google TPU and MIT Eyeriss. Results validate the advantages of domain-specific, reconfigurable architectures for multifunction RF systems, combining high performance, scalability, and energy efficiency.
Accurate modelling of propeller performance is essential for predicting thrust, torque, and efficiency across diverse operating conditions, enabling the design of efficient and reliable rotor systems. This study presents a hybrid modelling approach that integrates Blade Element Momentum Theory (BEMT) with Vortex Theory to improve performance predictions of amphibious propellers operating in dual-media environments, namely air and water. Classical BEMT, while widely used, suffers from inherent limitations such as inaccurate handling of high axial induction factors, stall behaviour, and tip losses, particularly under transitional flow regimes. To address these shortcomings, the study proposes a modified BEMT framework incorporating high-induction factor corrections by Shen and Spera, along with the Viterna–Corrigan stall model to account for post-stall airfoil behaviour. In addition, vortex-based modelling is employed to capture tip vortex dynamics, which classical BEMT fails to represent accurately. The hybrid model combines sectional blade discretization with velocity-based transition criteria to determine the onset of stall and dynamic effects. The methodology is validated through comparison against an extensive set of experimental data from APC propellers operating under various conditions, including multiple propeller diameters, rotational speeds, and fluid velocities. Results indicate that the hybrid model significantly outperforms both pure BEMT and vortex-only models, especially in predicting performance across transitional regimes where air-water interaction dominates. Detailed error analysis shows a substantial reduction in average and maximum prediction errors for thrust and torque coefficients. The study concludes that the hybrid modelling approach provides a more reliable framework for amphibious propeller design, particularly in applications involving marine and aerial robotics. Future work is proposed to further refine cross-medium transition models and expand the framework to fully unsteady and three-dimensional flow conditions.
In recent years, multi-agent reinforcement learning (MARL) has emerged as a promising approach for multi-unmanned combat aerial vehicle (UCAV) autonomous countermeasure systems. However, conventional end-to-end MARL methods often lack expert knowledge guidance, leading to low training efficiency, which poses challenges for simulation-to-reality (sim2real) transition. In this study, we focus on the scenario of cooperative beyond-visual-range (BVR) aerial engagement involving multiple unmanned combat aerial vehicles. To address these challenges, we propose the hybrid-constrained multi-agent proximal policy optimization (HC-MAPPO) algorithm. First, we design a rule filter mechanism, where expert rules dictate agent behavior in well-understood states to ensure predictable and interpretable maneuvers, while the neural policy is applied otherwise. Second, we formulate the multi-agent aerial combat problem as a Constrained Markov Decision Process (CMDP) and incorporate a cost-critic network into the actor-critic architecture, which enables explicit estimation of long-term constraint costs and decouples penalty from task rewards. Third, we develop a bilevel optimization framework for constrained policy search, which provides theoretical convergence guarantees and demonstrates improved training stability over traditional Lagrangian-based methods. Empirical results demonstrate that HC-MAPPO achieves a superior success rate, improving the win rate by approximately 20\%–30\% compared to existing MARL baselines such as MAPPO and HAPPO. Ablation studies further confirm the necessity of both constraints: removing either one leads to performance degradation.
The use of rotary-wing Unmanned Aerial Vehicles (UAVs) to spray pesticides and other farming products offers significant benefits to agricultural practices. These systems offer advantages through their ability to target specific areas with different application rates, reduce chemical waste, and adapt to different field conditions. They can also be used to carry out supplementary tasks such as field mapping. However, UAV performance in this setting is dependent on design choices that affect stability and maneuverability. This then has an influence on the efficiency of the sprayer system as rotor-induced airflow can disrupt droplet dispersion, making the pesticide less effective. This research uses simulation and experimentation to evaluate how changes in the design of an agricultural UAV can affect and ultimately improve its performance. The project began by conducting a literature review, which guided the simulation and experimentation setup and identified the technical requirements of UAVs integrated with sprayer systems. This research also highlighted the benefits of implementing UAV sprayer systems in agriculture. Based on this analysis, it was determined that an effective UAV would have a hexacopter frame due to the lift capacity and stability it provided. A sprayer system capable of distributing water or pesticides would then be designed to be integrated with the UAV, and computational simulations could be run to model the rotor-induced airflow and predict the distribution. This could then be compared to experimental results and inform future design changes. The design of the UAV was divided into two systems, the UAV system and the attached sprayer system. The two were developed alongside each other to ensure that the sprayer system was structurally compatible with that of the UAV. However, the sprayer system components were finalized after those of the UAV system, which is currently being assembled and is nearly flight-ready, to ensure proper integration. The UAV design features a carbon fiber frame because of the higher strength-to-weight ratio it provides, allowing it to support a payload of 1 L and an estimated maximum takeoff weight of 6kg. The other UAV system components, including the motor, ESCs, flight controller, and battery, were chosen with the goal of maintaining a thrust-to-weight ratio of 2 and a flight time of 10 minutes while ensuring their electrical compatibility. Future work will involve integrating the sprayer system into the UAV frame to enable flight testing and allow for data collection to begin. The experimentation will primarily revolve around measuring the uniformity of the sprayer system under different flight conditions, with changes being made to the setup of the UAV depending on how the results compare to those of the simulations. This will provide insight into how the characteristics of the system can affect UAV performance and inform the best practices for their implementation in agriculture, demonstrating their value in agricultural applications.
Before 2025, no open-source system existed that could learn Lyapunov stability certificates directly from noisy, real-world flight data. No tool could answer the critical question: is this controller still stabilizable—especially when its closed-loop system is a total black box. We broke that boundary. This year, we released the first-ever open-source framework that can learn Lyapunov functions from trajectory data under realistic, noise-corrupted conditions. Unlike statistical anomaly detectors, our method does not merely flag deviations—it directly determines whether the system can still be proven stable. Applied to public data from the 2024 SAS severe turbulence incident, our method revealed that, within just 60 seconds of the aircraft's descent becoming abnormal, no Lyapunov function could be constructed to certify system stability. Moreover, this is the first known data-driven stability-theoretic method ever applied to a civil airliner accident. And our approach works with zero access to the controller logic—a breakthrough for commercial aircraft where control laws are proprietary and opaque.
Hypersonic flight enabled by scramjet propulsion has the potential to revolutionize commercial aviation and defense systems, yet the computational expense of high-fidelity design tools remains a significant barrier to widespread development. Computational Fluid Dynamics (CFD) simulations require 3-4 hours per design iteration and expensive software licenses, limiting accessibility for researchers, educators, and small-scale developers. This study presents the development of a machine learning-based reduced-order model using Random Forest regression to enable rapid scramjet performance prediction without requiring CFD expertise or computational infrastructure. The model was trained on 250 high-fidelity Ansys Fluent simulations of 2D scramjet geometries sampled via Latin Hypercube Sampling across three design variables: strut wedge angle (12-24°), combustion chamber height (6-14 cm), and intake ramp length (40-80 cm). Physics-informed feature engineering transformed raw geometric parameters into 16 features incorporating trigonometric functions, interaction terms, and oblique shock theory. Performance efficiency was quantified through entropy generation analysis, integrating mass-weighted entropy flux across control volume boundaries following second-law thermodynamics. Overall, the Random Forest ensemble achieved 4.73% mean absolute percentage error, 0.126 average root mean square error, and 0.025 mean squared error across five key outputs validated through 5-fold cross-validation, demonstrating 19.8% improvement over linear regression baselines. The trained model reduces prediction time from hours to 4.6 milliseconds while maintaining 92-100% of predictions within ±5% engineering tolerance. Paired with a physics-informed visual ROM that blends machine learning predictions with analytical shock relations, the framework provides instantaneous numerical outputs and Mach contour visualization comparable to CFD simulations. By democratizing access to scramjet design tools and eliminating expensive simulation licenses, this approach accelerates optimization workflows for academic institutions, aerospace companies, and independent researchers.
This research presents a comprehensive computational platform for model rocket performance analysis, integrating multiple advanced analytical modules to address critical aspects of rocket design and flight dynamics. The platform implements trajectory analysis using Newton's laws of motion with atmospheric density interpolation, Monte Carlo simulation methodology for probabilistic landing area analysis, and design optimization utilizing the Nelder-Mead algorithm for multi-parameter optimization. The system incorporates six-degrees-of-freedom analysis, thermal heating calculations, composite material property computations, and drag coefficient estimations. Static stability analysis employs Mach-dependent aerodynamic coefficient calculations to determine stability margins and neutral point positions. Built on Angular framework with TypeScript, the platform provides real-time visualization through Chart.js integration and generates comprehensive PDF reports for each analysis module. The platform's modular architecture enables independent analysis of aerodynamic, structural, and performance parameters, facilitating evidence-based design decisions for model rocket development. Validation studies demonstrate an overall accuracy of 87.2% across all computational modules, with trajectory analysis showing the highest precision. This computational tool represents a significant advancement in accessible rocket science education and research, providing professional-grade analysis capabilities previously limited to specialized aerospace software. The platform serves as an educational resource for students and researchers while offering practical design optimization tools for amateur rocket enthusiasts and engineering professionals.
The Distributed Object Manager (DOM) is a general-purpose, extensible, customizable, high-performance, object-oriented cataloging system that is used for data operations by dozens of flight missions at NASA’s Jet Propulsion Laboratory. DOM has been actively supporting missions since its inception in 1993, when it was first developed to support JPL space mission operations. DOM's lasting relevance provides a unique context from which to draw valuable lessons in software architecture, maintenance, and real time operations as JPL looks forward to the next generation of multi-mission data management systems. DOM predates the widespread adoption of modern cloud-based data management technologies by more than a decade, while providing much of the same functionality. DOM empowers missions to define custom schemas and object types, manage user permissions and privileges, and customize notification services; this allows DOM to meet mission-specific needs throughout entire mission lifecycles, many of which span multiple decades. Additionally, DOM features lightweight, distributable catalog servers and provides command line interface (CLI) tools and graphical user interface (GUI) clients to access the different object store servers enabling a full suite of create, read, update, and delete (CRUD) database operations. DOM continues to support critical uplink and downlink data transactions via the Deep Space Network for many flagship missions from Voyager through Europa Clipper. Over the past three decades, DOM has undergone several infrastructure modernizations while maintaining its core functionality and effectiveness. DOM has faced various technical and operational challenges, such as scaling to support an increasing number of flight projects, extending to support growing mission data volumes, coordinating maintenance downtime, ensuring consistency across significant administrative personnel changes, and incorporating emerging technologies and best practices. To address these, DOM’s core functionality has been extended to include a remote method invocation (RMI) interface, a file notification service (FNS), event-driven client programs (Message Reactors), and more. We present insights from the current DOM administrative team on evidence-based operations and data management best practices. DOM's fundamental challenge has been to provide continuous mission support while adapting to changes in data, hardware, cybersecurity, and mission requirements. In adapting to these challenges, DOM exemplifies a successful approach for administering a mature, lasting, reliable flight mission data management system.
The rate at which satellites are being placed in orbit is creating visible challenges on the management, development and maintenance of ground segment software applications. Ensuring maintainability and reliability, while scaling systems and teams is a challenge, particularly across a diverse set of specialised domains such as the satellite operations sub-segment with its critical applications such as mission control, flight dynamics, mission planning, simulation, among others. In the context of Mission Control Applications and Tools maintenance and development, EUMETSAT has defined a common Application Lifecycle Management Requirements (ALM-R) for Mission Control Applications and Tools (MCAT) covering engineering and processes specifications aimed to ensure that lifecycle of any/most of mission control applications is defined in a cohesive, coherent and-harmonized way. The requirements have then been mapped to an engineering Application Lifecycle Management System (ALM-S) made up of tools and implementation processes enabling the different stakeholders (e.g. developer, verification manager, librarian, QA team) to manage the phases of the engineering lifecycle according to their role and responsibility. The current combined ALM-R and ALM-S is the result of years of lifecycle management evolution and lessons learned which has been so far considered reasonably fulfilling the overall goal, but recently showing difficulties to scale with the increase number of ground segments to develop and maintain, complexity in terms of knowledge (technologies, design, etc..) and processes management, effective and intelligent automation, all projecting teams in the future to act with a suboptimal level of efficiency. In addition, this scenario is experiencing increasing budget constraints raising concerns on efficiency sustainability. Due to all these factors, assessment towards an ALM-R and ALM-S evolution addressing the mentioned factors has become a clear strategic necessity which we have addressed in the definition of a roadmap including indeed evolution, but also new technologies taking advantage of Artificial Intelligent (AI) technologies and paradigms. AI has been identified having a core role, specifically deployed with assistant and agentic capabilities to equip stakeholders with a configurable AI + Human-in-the-loop integration addressed by a system design capable to evolve both in terms of use-cases and technologies advancements. The approach covers both a revision of known ALM use-cases with their formalization into use-case taxonomy and in parallel their mapping to an advanced prototype implementation of a system layered architecture. The layered architecture presents at its bottom a knowledge base layer in charge to store and maintain into vector databases information artefacts from different data sources and made them available to the upper agentic layer (in charge of implementing back-end use-cases functionality) and the presentation layer providing ALM end-users with diverse UI/UX access to ALM use-cases. This paper introduces what we consider the fundamental phase of the roadmap for the ALM-R and ALM-S evolution named AI Augmented-ALM (AA-ALM) and specifically its implementation solution referred as AA-ALM-System (AA-ALM-S).
Anomaly prediction is a critical method used by aerospace engineers to ensure overall reliability, safety, and operational efficiency of space missions. The proposed model takes advantage of deep learning approach that combines autoencoders, attention mechanisms, and recurrent neural networks to detect unusual patterns in time series data from space missions. This allows the model to identify both spatial and temporal anomalies within the data. A key innovation is a new range calibration mechanism inspired by alpha-beta pruning. This approach dynamically detects the anomalies based on threshold values and assists in reducing false positives within feature ranges. Additionally, the model also employs a series of sequential post-processing techniques to optimization the overall F1-score for anomaly prediction. The proposed approach is evaluated on two real-world datasets Soil Moisture Active Passive (SMAP) and the Mars Science Laboratory (MSL) rover (Curiosity) from NASA. Performance results demonstrate superior performance when compared to the other baseline models. The approach can proactively mitigate mission-critical system failures, efficient resource allocation, and mission success in aerospace systems.
This work presents a generalized actuator modeling approach that enhances simulation accuracy, with applications extending beyond Mars exploration. Actuators are integral components in space robotics, where precision and reliability are paramount. The circa-2023 Sample Retrieval Lander (SRL) design exemplifies this, relying on a robotic arm to retrieve Mars sample tubes for return to Earth. With Earth–Mars round-trip communication delays on the order of minutes, direct human teleoperation is unsafe and impractical, necessitating autonomous execution through onboard kinematic controllers. Because the robotic arm represents a single point of failure in such a critical mission, we develop an exceptionally high-fidelity simulation environment to validate the controller design. Unlike many system simulations that idealize actuator performance to simplify computations, our model captures key dynamics often overlooked in traditional approaches. Its accuracy is validated through comparisons with real-world actuator tests, demonstrating both mission relevance for SRL and potential applicability to future space robotics systems.
This study presents a systematic approach to identifying the most suitable mining regions on the Moon. The developed method simultaneously evaluates various data layers to determine the most favorable areas for potential in-situ resource utilization (ISRU). The analysis integrates geochemical data such as iron (Fe), titanium (Ti), and thorium (Th); topographic variables like surface slope and elevation; and environmental factors including sunlight duration and regolith structure. The Analytic Hierarchy Process (AHP) method was used to assign values to each parameter according to its importance for lunar mining operations. The goal is to strike a balance between scientific potential and engineering feasibility. Locations were ranked based on composite suitability scores, enabling prioritization of regions that are both resource-rich and operationally accessible. The multilayered analysis allows for a more detailed and balanced spatial evaluation of the lunar surface, revealing areas where advantages intersect. The study adopts a flexible decision-making model that can adapt to various mission scenarios, from robotic explorations to crewed missions. The AHP method enables an objective assessment of the importance of criteria, reducing subjectivity in site prioritization. Preliminary results indicate that a spatial trade-off is required between high mineral content and challenging terrain conditions. This method has facilitated the creation of detailed suitability maps due to the abundance of resource data. The validity of the model was tested by comparing it with known geological features and mission planning documents. Furthermore, the resolution limitations of orbital data were addressed, and strategies were proposed to reduce spatial uncertainties in surface characterization. This research contributes to the field of space resource utilization by providing: (1) a transparent and reproducible decision support tool adaptable to different mining objectives; (2) a quantitative evaluation of the effect of parameter weights on spatial prioritization; and (3) foundational insights for future in-situ validation missions. Although the proposed system focuses on lunar applications, the methodology can be applied to Mars or asteroid exploration missions with appropriate parameter adjustments. In conclusion, integrated approaches that systematically assess both technical and economic factors can significantly improve pre-mission planning and reduce risks for future commercial and scientific ventures in space.
Reinforcement learning (RL) has emerged as a popular choice for training artificial intelligence algorithms in competitive scenarios due to recent successes of achieving superhuman performance in board and video game environments. However, there is currently a lack of open-source, competitive environments designed to exhibit challenges of realism in the domain of spacecraft control to assist in the transfer of RL algorithms to physical systems. We present AstroCraft, a space-based capture-the-flag environment comprised of two opposing teams, each containing multiple maneuverable satellites and a space station in geosynchronous equatorial orbit. The primary goal of each team is to maneuver a satellite to capture a flag at the opposing player’s station and return the flag to its own station without being tagged-out by opposing satellites. We first perform an experimental study on AstroCraft that elicits challenges from realisms such as long time horizons, complex dynamics, and stringent fuel constraints. Throughout, we conduct experiments on similar environments from the literature indicating that many of these realisms are not significantly present and do not degrade performance. Finally, we design an RL algorithm which first gathers data from a heuristic opponent competing against itself; constructing this dataset enables the application of Conservative Q-learning for offline pretraining before further online finetuning. This algorithm produces a model that is superior to the original heuristic opponent. We believe that the lessons learned from our experiments on AstroCraft provide promising avenues for constructing RL algorithms that overcome challenges of realism in simulation and physical systems alike.
Accurate and robust pose estimation of non-cooperative spacecraft is critical for autonomous rendezvous and on-orbit servicing. While monocular vision-based methods have attracted growing interest owing to their low cost and structural simplicity, achieving high-precision pose estimation under large scale variations in target distance and complex illumination conditions remains a formidable challenge. In this paper, we propose a novel dual-path prediction network reinforced with a geometric consistency constraint to address these issues. Our framework features two distinct yet complementary pathways. The first path employs a feature pyramid network to extract multi-resolution representations, from which stable keypoints are detected and subsequently integrated with a PnP solver, thereby enabling accurate pose estimation across targets with large scale variations. The second path employs an adaptive-weighted feature pyramid network augmented with a spatial self-attention module to effectively fuse multi-scale information and strengthen global contextual reasoning. Its output is processed by two direct regression heads for rotation and translation, hence improving accuracy and robustness under occlusion and degraded geometric conditions. To ensure coherence between the two pathways, we further introduce a geometric consistency loss that enforces alignment of their outputs during training, thereby improving stability and generalization. Experimental results on SPEED and SwissCube datasets demonstrate that our framework achieves substantial improvements over existing methods, particularly under extreme conditions.
In order to maintain the system performance and mission productivity of spacecraft, health monitoring and fault diagnostics are crucial. It's essential to ensure that a spacecraft is operating properly without anomalies, as it could jeopardize the whole mission. Traditional approaches are challenging to apply, as they often rely on post-mission data or resource limitations, which are insufficient for detecting subtle or emerging anomalies during flight. This paper introduces a simulation-driven diagnostics framework that uses NASA's SimuPy and a custom telemetry generator to emulate fault conditions and produce multivariate telemetry streams for spacecraft. We apply Random Forest classification models to the synthetic telemetry to detect and categorize anomalies. The proposed framework facilitates both pre-launch validation and in-flight anomaly detection. While previous approaches have primarily focused on retrospective failure analysis, our approach supports proactive diagnostics by simulating system behavior and dynamically injecting both nominal and fault conditions during runtime.
Fatigue crack growth under high-cycle fatigue is one of the most severe problems in the design, maintenance, and safe operation of aircraft structures. During operation, these structures experience millions of loading cycles, which cause the gradual growth of nucleated cracks leading to ultimate failure. Thus, accurate modeling of fatigue crack growth is necessary for ensuring structural integrity, maximizing inspection intervals, and extending the service life of aerospace structures. Conventional methods of modeling crack growth under high-cycle fatigue use linear elastic fracture mechanics to arrive at the cyclic stress intensity factor, which is then used in the Paris Law describing steady state crack growth. Paris law is highly non-linear, consisting of two constants, C and m, under fully reversible loading. These parameters are evaluated using Euler integration of Paris law and linear regression of scattered crack growth measurements from standard tests. However, a lack of constraints during the calibration can render the parameter estimates to be inaccurate particularly when the data is significantly scattered. To address this limitation, Physics-Informed Machine Learning (PIML) architectures are employed to calibrate the parameters of Paris law. However, before utilizing this calibration approach, the accuracy of Physics-Informed Neural Networks (PINNs) to integrate Paris law was tested. To this end, the predictions from Physics-Infused Long-Short Term Memory (PI-LSTM) and Implicit Euler Transfer Learning (IETL) architecture were also compared to Euler integration, and a reasonable agreement was obtained. Following this, these methods were applied to obtain the parameters from numerically generated data using some assumed C and m values. It was observed from the study that the method was not only able to calibrate the parameters, but also that the network could be used to predict crack growth when the amplitudes were modified. Finally, scattered data was artificially generated by choosing distributions of C and m. Subsequently, IETL was applied to the scattered data to calibrate the parameters and showed a satisfactory comparison. In summary, this study exemplifies the merits and demerits of different PIML methods when applied to predict crack growth from the Paris law. Furthermore, the approach allows both crack growth evolution and Paris constants to be predicted from limited experimental data, thereby reducing the need for repeated costly tests across different loading cases. Finally, the reliability of the PIML framework to predict crack growth for various amplitudes and block loading is demonstrated.
Aviation maintenance is a foundational pillar of operational safety, reliability, and mission success. As aircraft systems evolve to incorporate increasingly sophisticated technologies such as advanced avionics to integrated sensor networks maintenance personnel face mounting complexity in diagnostics, data interpretation, and procedural execution. These challenges are compounded by workforce shortages, Subject Matter Expert (SME) turnover, and the growing demand for predictive, data-driven maintenance solutions. In particular, the departure of experienced personnel often results in the loss of critical experience, including nuanced troubleshooting strategies and platform-specific insights that will challenge the ability to sustain operational continuity or support the next generation of maintainers. The traditional documentation and informal tribal knowledge, while historically effective, are increasingly inadequate in the face of complex systems, workforce turnover, and the demand for predictive insights and scalability. In both commercial and military domains, these limitations contribute to reduced readiness, inefficiencies, and elevated lifecycle costs. Recent advancements in Artificial Intelligence—particularly Large Language Models (LLMs) and predictive analytics—present a novel but complex opportunity to address these challenges. LLMs can ingest and synthesize unstructured data from manuals, technician notes, and maintenance logs, enabling intelligent diagnostics, contextual troubleshooting, and dynamic knowledge sharing. This paper will examine traditional maintenance practices and evaluate how different AI models can mitigate systemic inefficiencies while critically addressing the operational, ethical, and technical constraints inherent to their deployment.
Mobile multirobot systems have become increasingly utilized due to benefits such as redundancy, increased coverage, and the ability to create improved data products (like using several scalar field measurements to compute an instantaneous gradient). Multirobot systems include a wide variety of architectures; ranging from swarms, where large numbers of relatively simple robots form loose formations, to more centralized architectures where relatively few robots move with tight formation control. An example of a more centralized architecture is cluster control, which has been developed by the Robotic Systems Laboratory at Santa Clara University to control formations on land, in the water, and in air. Real world testing of multirobot systems can be challenging, therefore, there is a need to develop a robust indoor testbed that is easy to use for experimental verification, reliability and data analysis. In order to create the testbed, we optimized an OptiTrack motion capture system to track Crazyflie 2.1 microdrones within an enclosed 3D space. To achieve this, twenty-four infrared cameras were positioned around the 6 m by 3 m by 3 m workspace for precise tracking in 3D, and thoroughly calibrated. The resulting position data had error margins below 2 cm. This real-time position data was then broadcast over ROS2 for closed-loop control of the microdrones via PID controllers. After a brief description of cluster control, a two drone cluster definition is presented and forward and inverse kinematics as well as the inverse Jacobian are developed. In our paper, we will present Optitrack position data, plots illustrating the flight performances of single drones, as well as results of two-drone cluster flights. Performance tests include hovering, step responses, and multi waypoint navigation. Individual drone performance will be compared to cluster performance, in order to address the differences cluster control architecture provides. Additionally, we will include position error results, and standard deviations of error illustrating the successful implementation of PID tuning and cluster control. These results provide further support for the implementation of cluster control architecture, which can then be used for future formation control experiments.
Space programs face budget cuts and cancellations as their benefits may not justify their cost. In other words, their value (here: benefits minus cost) is insufficient or has not been identified (e.g. scientific gains, job creation). Defining the potential value of space programs is best addressed during its conception, i.e. architecting phase. Space program architecting approaches from literature do not explicitly consider the link between the system architecture and value delivery. We propose to systematically identify ways how value is delivered by a space program architecture from proven value delivery mechanisms. Those proven value delivery mechanisms are captured in the form of value creation patterns. Patterns capture problem-solution knowledge for a specific context. They were first introduced in architecture and later popularized in object-oriented software engineering. They were further applied in systems architecting, and recently in space systems architecting. We first develop a conceptual data model of space programs to structure organizational and technical concepts relevant to space programs, and the relationships between them. This is grounded in the ECSS Glossary and the NASA Systems Engineering Handbook. We then build a database of preliminary value creation patterns in space programs. Examples include a “dual use” pattern that was sourced from a review of the Luxembourg space sector, where the context is that the country’s space policy seeks to employ space infrastructures for the benefit of sectors other than space. The problem there is how value can be created for other sectors by using space infrastructures. Factors influencing the solution include development cost and commonality. One solution is to develop systems for dual use in space and on Earth. An example is the Luxembourg company Maana Electric, which develops ISRU appliances which can produce solar panels from sand on Earth and regolith on the Moon. Another example is the “diffusion” pattern. The context is a country with a non-space-related industrial base. The problem is how to advance the state of the art in that industrial base while contributing to space system development. Similarly to the “dual use” pattern, a key factor is the architectural similarity between the terrestrial and space systems that are developed. The solution is to utilize the capabilities of that industrial base in the development of a space system. A historical example is the Canadian STEAR program, where the country’s robotics industrial capability was applied in the development of the ISS Mobile Servicing System. To explore many similar patterns, complementing manual search, we use a Large Language Model (LLM). This LLM is then used to semantically search through the NASA Technical Reports Server and the ESA Data Discovery Portal for patterns matching or resembling the preliminary value creation patterns. This approach precedes a trade space exploration where space program architectures are designed using patterns, given a certain definition of value that may vary from different actors’ viewpoints.
Nowadays, successfully completing a strategic project is essential to ensuring an organization’s survival. This holds true for most companies whose goal is to sustain their operations and expand their market. The aerospace and aeronautical manufacturing sectors are undergoing a profound transformation driven by the accelerated integration of digital tools and innovative technologies. These changes are reshaping traditional project management practices, especially in the context of complex engineering projects. The adoption of IT tools, planning management software, 3D printing, computer simulation, or AI are examples of technologies that reduce the need for human and capital resources while simultaneously increasing design efficiency. This research focuses on the relationship between the use of various technologies in project management (especially R&D, design, and continuous improvement projects), the methodologies employed, and the performance indicators that measure and control the factors defining project success. This research adopts a qualitative approach based on semi-structured interviews with 25 professionals from the aerospace and aeronautical sectors across Canada, the United States, and France. The participants include project managers, engineers, and digital transformation specialists from leading organizations in civil aviation, major aerospace firms, and key systems integrators. The data were analyzed using thematic coding to identify recurring patterns and insights. Core topics of the discussions included the use of digital technologies, the application of project management methodologies (traditional, agile, or hybrid), and the perceived impact of these elements on project performance. These themes served as the analytical backbone of the study and guided the interpretation of results. The findings indicate that the adoption of digital tools positively correlates with project performance - particularly in terms of schedule adherence, cost control, and risk mitigation - when such tools are embedded within a coherent project governance structure. Metrics such as the Schedule Performance Index (SPI) and Cost Performance Index (CPI) were commonly used to monitor progress. However, advanced technologies like virtual reality and digital twins are not yet widely deployed across the sector. In contrast, solutions such as 3D printing, computational simulation, and project management software are more broadly adopted and integrated into daily operations. Artificial intelligence (AI) is an emerging trend, showing strong potential, yet its adoption remains constrained due to concerns over data sensitivity and a frequent reliance on in-house development. Moreover, the study reveals that the most effective outcomes are observed when organizations adopt a hybrid project management methodology - blending agile and traditional approaches - combined with simple, well-integrated digital tools tailored to the project context. This study contributes to a better understanding of the interplay between digital transformation and engineering project performance. It underscores the importance of adopting a systemic and strategic perspective when implementing digital solutions. The results offer actionable insights for decision-makers aiming to align technology investments with performance objectives. By shedding light on the enabling and limiting conditions of digital integration, this research helps bridge the gap between technological promise and practical impact in aerospace project environments.
Projects in NASA’s Pre-Phase A aim to define a mission concept that is aligned with program goals, technically feasible, and affordable enough to merit continued development. However, this work is conducted under tight constraints—limited time, budget, and workforce—despite the need to generate, assess, and compare a wide range of mission concepts. Without structured tools to guide this process, this early concept formulation can become inefficient, increasing uncertainty and risk in later phases of mission development. Traditional concept formulation relies on discriminative tools—models that use features of a mission concept (e.g., mass, power) to predict target outcomes such as financial viability (i.e., cost). While effective for evaluating specific designs, these tools do not efficiently support the generation, assessment, nor comparison of a broad range of alternatives. Generative models are a promising alternative for overcoming these limitations. Rather than predicting outcomes from features, generative models identify the accessible design space of features from the desired target outcomes such as technical feasibility and financial viability. By guiding concept formulation to exploration within an informed population rather than by brute force or random trial-and-error search methods, generative models significantly reduce the time, labor, and finances required to identify selectable mission concepts. This paper presents the development and application of generative models at NASA’s Jet Propulsion Laboratory to support early-stage mission concept formulation. These results demonstrate the potential of generative models to transform early mission formulation into a more strategic, resource-efficient, and selection-ready process.
