Past and Present: from a figurative “rocket science” to omnipresence
When I linked from my blog to chief Saab testpilot Marcus Wandt’s Gripen/E Software-First video , I wasn’t sure he was Marcus-the-astronaut. The blog was about Edge computing & ML; he’s a Chalmers Tech graduate in electronics & AI and mentions ML from sensor-fusion data (multimodality). As an apology for omitting his work for ESA in space, let’s bring out AI’s contribution over the past 25 years to aerospace.
In the early Space Shuttle era of 1990´s, I presented at two AI congresses sponsored by NASA and Kennedy Space Center (KSC). Shuttles have changed completely since then, and so has applied AI.
The critical steps of a mission were takeoff, docking, spacewalks, and last (but far from least) landings. The descent under a precise angle of attack got just enough floatability (buoyancy) from body and small wings of the spacecraft; an anecdotal sum-up was: build a model using a brick as a body and cardboard as wings – and make it fly. The Shuttle Project therefore used the perfectly even 4.6 km runway at KSC, FL, as its prime landing site (78 landings). Edwards, CA (59) and White Sands Space Harbor, NM (1), were backup sites. Along with that, intelligent avionics for safe takeoffs and landings became crucial. Cognitive “programmer-made” AI ruled (pun intended) in the 1990´s. Most apps used rule-based inferencing (rulebases and rule engines).
A Space Shuttle returning to KSC “piggybacked” on a B747 JumboJet (source: Wikipedia.org)
In contrast today, AI is part of the aerospace DNA: onboard electronics, mission control, fault prevention, robotic missions, in-space production, HMI/visualization, sensor fusion, signal filtering, and dozens more.
NASA’s long priority venture in open data/open source/open science on astrobiology is especially interesting, now that deep ML has revolutionized life sciences, ranging from diagnostics, protein folding and mRNA vaccine to rehabilitation, body monitoring, injury prevention, fast worldwide R&D data exchange using federated ML to remote-surgery platforms. NASA hopes to boost the search for life on other planets and moons, as well as ways to train and safeguard human crews on extremely long missions with exposure to conditions such as vibrations, radiation, or weightlessness. Again, one example among a lot.
The federal FAA and its EU partner EASA also take a positive interest in trustworthy AI and making it human-centric.
Hopefully, this will compensate for my miss on Marcus Wandt’s ESA work. It’s time for a structured sum-up; years back it would take quite many hours, but today we use generative AI instead. Here we go, Microsoft Bing:
Copilot: In summary, AI and ML have become integral to modern spacecraft, enabling autonomous navigation, data analysis, and decision-making. The shift from rule-based systems to data-driven approaches has revolutionized space exploration and mission efficiency. AI and ML have significantly evolved in the context of spacecraft from the late 1990s to the 2020s.
Key differences:
1990s AI:
AI techniques primarily rule-based, symbolic. Handcrafted knowledge bases and inference engines, expert systems for tasks like fault detection, diagnosis, and decision-making.
Limited computational resources, lack of scalability.
1990s ML:
ML algorithms existed, less prevalent. Supervised learning (e.g., regression, classification) for specific tasks. Feature engineering was manual and required domain expertise.
Data availability and performance limited.
2020s AI:
More data-driven and statistical approaches. Deep learning dominates, especially in computer vision and NLP. NN used for image analysis, anomaly detection, autonomous navigation.
Reinforcement learning (RL) is explored for spacecraft control and optimization.
2020s ML:
ML pervasive.
Supervised learning: Improved models, larger datasets, better algorithms. Unsupervised learning: Clustering, dimensionality reduction, anomaly detection. Semi-supervised learning: Combining labeled and unlabeled data.
Transfer learning: Pretrained models fine-tuned for specific tasks.
AutoML: Automated model selection, hyperparameter tuning, and feature engineering.
Edge ML: Onboard ML processing to reduce communication latency. Quantum ML: Exploring quantum computing for complex simulations.
More:
1 en.wikipedia.org, 2 esa.int, 3 hyperight.com, 4 pinterest.se, 5 northerngrafics.se, 6pinterest.se (plus blogs)