Leveraging Human Intelligence in AI/ML Approaches for Space
From Human Mind to Machine
With the nature of our industry being in software, our company has had the opportunity to be part of the revolution that is Artificial Intelligence (AI) and Machine Learning (ML). While AI relates to the idea that machines can mimic how a human acts when given certain inputs, ML aims to teach machines how to respond to inputs and give accurate results based primarily on pattern recognition. The power of ML lies in its extraction of patterns from training data (or examples that it is given) and its ability to act accordingly. There are currently two main ideologies behind AI/ML: model-centric and data-centric. In a traditional academia setting, students are often given pre-cleaned data where they can work to find the best model to fit the dataset. This approach is a model-centric approach that is beneficial for teaching new AI practitioners various ML models such as clustering, decision trees, classifiers, neural networks, etc. However, in the real world, data available for training is not perfect and must be cleaned and processed before use. This leads into the data-centric AI approach, which is the discipline of systematically engineering data to create better AI models. Data-centric methodology makes use of available processes, including AI algorithms, to focus on improving the quality of data to drive ML training optimization. Other data-centric techniques include augmentation to supplement limited truth data with unlabeled or synthetic data.
At Millennial Software, we lean towards the data-centric approach with recognition that real world data needs to be processed and augmented before being useful, but also with the understanding that human expert knowledge is critical. Central to AI/ML is human intelligence, which guides how both model-centric and data-centric systems function. From this perspective, rather than simply processing data and producing outputs based only on mathematical and algorithmic models, our AI/ML approach involves prioritizing human intelligence first. This methodology calls for human expert knowledge to drive feature engineering and data pre-processing before ML training to result in effective solutions. As a team of over forty specialized engineers, Millennial Software has leveraged our human-centric AI/ML approach in competitions like the MIT ARC Lab Prize for AI Innovation in Space, accelerators like the SDA TAP Lab’s Apollo Accelerator, and as part of our customer solutions.
Augmenting AI/ML for Space Solutions
What happens when AI/ML is applied to the space domain? The reality is that its potential is vast. AI/ML can augment space personnel, solutions, and missions, extending our ability to explore, understand, and interact in space. Furthermore, innovative AI/ML solutions can enable us to tackle the most complex challenges faced in Space Domain Awareness (SDA) by improving data processing, prediction accuracy, and operational efficiency. However, in our experience, fine-tuning and improving AI/ML model performance for complex space challenges often requires a human touch. This is where the integration of human intelligence is valuable. To augment AI/ML, we can utilize feature engineering that incorporates human expert knowledge of topics like orbital phenomenology, historical and predicted threat behavior, and black swan event designs.
Millennial Software has shown how using a human-centric approach can effectively augment datasets for successful training of AI models, as seen in our model performance results for the MIT ARC Lab Prize for AI Innovation in Space. For that challenge, we leveraged our extensive experience in astrodynamics and computer science to develop AI algorithms to track and predict satellites’ patterns of life in orbit using passively collected data. Our process was iterative and involved encoding our logic into feature engineering and harnessing human intelligence to teach ML models to extract necessary patterns from data. Our solution and approach was presented at the 2024 Advanced Maui Optical and Space Surveillance Technologies (AMOS) Conference.
Millennial Software team at the 2024 Advanced Maui Optical and Space Surveillance Technologies (AMOS) Conference.
When it comes to building predictive systems and applications, we take the same human-centric approach to support SDA challenges. As subject matter experts in SDA and astrodynamics, we understand that truth data needed for AI/ML training is severely lacking. Using models to generate training data is simply not possible due to the nuances of the SDA context, which are often scenario and behavior based. For example, indicators of camouflage, concealment, deception or maneuver (CCDM) are based on threat behaviors that cannot be modeled by AI for the purpose of data generation. Similar to individual ML model training but on a much bigger scale, SDA systems suffer from even more scarce data. This is especially true in our experience with the CCDM context at the SDA TAP Lab. In order to build predictive models, we must first build the initial prototype system, use it to generate human verified truth data, and catalog all inputs and outcomes. The new truth dataset can then inform synthetic data generation through scenario creation and simulation. In aggregate, the augmented dataset is driven by actual operational context and is most relevant to SDA battle scenarios, making it the perfect starting point for AI development. This culmination can result in a system that generates relevant data to improve AI models for predictive use.
The Impact of Data
Understanding AI/ML solutions requires a closer look at the pivotal role that data plays. Before ML training, and even before data processing and augmentation, there needs to be a relevant and varied dataset. In our recent blog post, From Ideas to Impact: Apollo Accelerator Cohort 4, we highlighted our Target Model Database (TMDB), which serves as a centralized repository for cataloging the key attributes of all space objects. This aggregation of space data from open-source and partnering companies has set the foundation for our work within the SDA TAP Lab. The impact of data can be seen in a recent application of our TMDB to identify a Russian launch Bar-M payload, which is generally associated with intelligence, reconnaissance, or military purposes. After successful payload identification, Millennial Software’s Coplanar Launch Predictor was used to estimate the delta plane and semi-major axis (SMA) adjustments needed to achieve orbital alignments. Combined, our tools provided valuable information to contextualize SDA.
Our TMDB has been a critical data source for our cohort’s applications, and shows how a dataset can serve current heuristics solutions while setting the stage for future AI/ML implementation. We see an opportunity to apply AI/ML to solve SDA challenges such as CCDM interrogation (detection and alerting) and direct-ascent anti-satellite (DA-ASAT) launch classification. Given the data made available by the TMDB, these challenges are primed for ML pattern extraction. For the Apollo Accelerator’s Cohort 5, Millennial Software is planning to expand the data store to include other aspects such as launch sites, launch vehicles, and events. When added to the existing datasets, they will serve as the all-encompassing source catalog of a Battle Management System (BMS). The BMS will deliver a software application to support SDA use cases such as interrogating CCDM behavior and preventing operational surprise. Once the BMS is operational at the end of next year, valuable information will be generated and verified by human experts and operators, which will produce the very data needed in the human-centric approach to building predictive systems. We are excited to continue our work in Cohort 5, where we hope to partner with other cohort members to leverage AI/ML augmented with human intelligence for our SDA solutions.
SDA TAP Lab Apollo Accelerator Cohort 5 members
Shaping Space Domain Awareness
As the world around us continues to transform, AI/ML will be a major driver towards the future of technology and evolution of tasks that would typically require a human. We already see the impact of AI/ML on space solutions, enhancing our ability to explore, understand, and responsibly shape our world. Now, leveraging human intelligence in these AI/ML approaches, we can incorporate human expertise and experience to build effective models and systems for problem-solving. It is clear that the challenges we see in SDA are particularly complex and nuanced. A human-centric approach to AI/ML has the potential to shape and enhance our space capabilities.
Millennial Software is proud to be at the forefront of tackling software-defined architecture challenges. Our contributions to the initial SDA TAP Lab BMS are setting the stage for a more dynamic and flexible defense infrastructure. By building a robust data foundation, we are laying the groundwork for the next wave of AI/ML innovations. In partnership with key industry players such as Quasar Satellite Technologies, Data Fusion and Neural Networks, and Planetary Systems AI, we are integrating advanced AI/ML capabilities that are augmented by human intelligence. This collaborative effort is pushing the boundaries of what’s possible in real-time decision-making, predictive analytics, and operational efficiency.
The future of AI/ML in the SDA ecosystem is bright. With our combined expertise, we are not just enhancing decision-making and situational awareness, but enabling a higher degree of autonomy and strategic foresight. This synergy between cutting-edge technology and human oversight will help us not only see further but also steward progress with purpose, ensuring our actions are aligned with long-term goals and national security objectives. Together, we’re shaping a future where intelligent systems not only respond to today's challenges but anticipate tomorrow's needs, transforming defense operations with agility and precision.