Terminal guidance laws: a comprehensive review from classical methods to reinforcement learning

Summary

A single-author survey unifies the entire spectrum of guidance laws for guided munitions—ranging from classical proportional navigation derivatives to GENEX, sliding mode, and differential game methods—under a common framework. The paper also positions reinforcement learning-based guidance as the latest extension of the dynamic programming tradition, characterized by zero effort deviation.

Why it matters

Guidance law design sits at the intersection of control theory, game theory, and increasingly, machine learning. The holistic comparison across information requirements, robustness, computational load, and theoretical guarantees provides engineers with a practical decision map. By framing classical and reinforcement learning methods as two ends of a single principle, the paper offers a useful perspective for teams evaluating hybrid architectures, rather than treating machine learning as a black-box replacement.

Highlights

• It is demonstrated that all methods (PN family, GENEX, sliding mode, HJI differential games, and reinforcement learning) share a zero effort deviation and dynamic programming backbone. This is the paper's central unifying claim.
• The comparison is structured around four concrete axes—information requirements, robustness, computational load, and theoretical guarantees—rather than an ambiguous list of pros and cons.
• The author argues that the field is moving toward hybrid designs that combine the analytical reliability of classical laws with the adaptability of learning-based controllers.

Critical take

This work is a single-author, non-peer-reviewed preprint that contains no experimental or simulation results, lacks visible citations in the summary, and has no download data yet. Consequently, the claim of being 'comprehensive' relies entirely on the author's own classification and cannot be independently verified. The assertion that reinforcement learning guidance is merely the latest step in the classical tradition carries a strong interpretive lens. This perspective risks downplaying the genuine architectural differences between learning-based and analytical controllers (e.g., distributional assumptions, sample efficiency, and safety certification).