systems theory//engineering patterns//technology maturity

2026-07-30

Technology maturity is a grading of each technique by how real it is in a given domain, on four levels: fundamentals (theory that does not expire, such as Bayes, Fourier, least squares and stability), industry (in production almost everywhere), niche (in production in specific places) and research (papers and demonstrations). It is used to decide what to build on, what to pilot and what to watch, and to read a vendor's pitch or a conference demo for what it actually proves.

The grade belongs to a technique in a domain, never to the technique alone. MPC has run refineries for decades and is research on a drone dodging obstacles at 20 m/s. The extended Kalman filter is industry in navigation (PX4's estimator is one), and the PID controller runs the large majority of industrial loops. Particle filters localize mobile robots in warehouses with a known map, a niche. A drone flown by a policy trained with reinforcement learning beat champion human pilots in 2023 (Kaufmann et al., Nature), real and impressive, on a known track, in controlled conditions and far from any certification. And the move between levels has no fixed timetable: the Kalman filter was published in 1960 and was navigating Apollo within the decade, while other ideas spend thirty years in conferences.

Three questions place a technique: who keeps it in production, since when, and with what data and what safety analysis? Is it in standard libraries, standards or manufacturers' manuals? Has it been reproduced outside its author's lab? If every answer is no, it is research, however good the video. A demonstration video is a sample of one, chosen by its author.

The frontier moves, almost always in one direction, from research toward industry. The book's map places event cameras, learned estimators with guarantees, reinforcement learning in critical systems and certified learned components in the research column today; ROS 2 in products and edge deployment of models in the niche; the microcontroller with an RTOS, PLCs, Kalman filters and gradient boosting on tabular data in industry.

Knowledge expires at different rates, by layer. Fundamentals (linear algebra, probability, optimization, differential equations, Fourier, graphs) last decades to centuries; methods (Kalman, PID, LQR, boosting, auctions, consensus) decades for the classics and years for fashionable architectures; systems (an autopilot, a fleet manager) years; tools (ROS 2, PyTorch, PX4, cloud services) years and sometimes months. The strategy follows: invest in the long half-life and learn tools just when they are needed, since an engineer who masters the fundamentals learns a new framework in a week.

Maturity is a cost input. A research-grade method brings unknown failure modes, no certification path and nobody to hire who has maintained it, which belongs in the total cost of ownership next to its accuracy.

The grade says nothing about whether a technique fits a problem; that is the flyswatter rule and context levels. Siblings in engineering patterns.