What Actually Works for Spacecraft Fault-Tolerant Control: An Honest Settled-Gate Benchmark of Learned and Classical Methods

Researchers benchmarked fault-tolerant control methods for spacecraft using rigorous testing criteria, finding that structured learning approaches combining gain estimation with analytic control laws significantly outperform classical and end-to-end learning methods on actuator faults, though constant bias faults remain unsolved without additional disturbance observers.

This research addresses a critical gap between simulation performance and real-world spacecraft reliability by establishing a demanding benchmark for fault-tolerant control systems. The study exposes a fundamental mismatch in the field: while recent learned methods report high success rates, those claims often rest on narrow test conditions and weak evaluation metrics. By requiring controllers to maintain pointing accuracy within 0.2 degrees across faults unseen during training, the authors create a realistic scenario that separates genuine robustness from statistical artifacts.

The finding that fault-unaware PID control and pure end-to-end reinforcement learning both achieve 0% success is particularly instructive. This demonstrates that raw learning capacity without domain knowledge fails completely—a counterintuitive result given the popularity of deep learning in control applications. The breakthrough comes from hybrid architectures that combine learned modules for online parameter estimation with classical analytic control laws. Structured designs achieving 97.8% success on gain faults versus 0% for unstructured approaches reveals that inductive biases matter more than model complexity.

The persistent failure on constant additive bias across all methods, including privileged oracles, identifies a fundamental theoretical limitation: integral-free control cannot reject constant disturbances. The paper's solution—adding a disturbance observer—recovers 59.4% of bias cases, suggesting practical gains remain possible through principled system composition.

For the aerospace and autonomous systems communities, this work provides both cautionary tales and constructive guidance. It shifts evaluation standards toward harder metrics that better predict deployment success and demonstrates that learned components work best when constrained by classical control structure rather than operating end-to-end. The released benchmark enables future comparison across heterogeneous methods.

• →Structured learning-classical hybrid controllers achieve 97.8% success on spacecraft gain faults while end-to-end RL methods score 0%, proving architecture matters more than raw model capacity.
• →Classical adaptive methods handle sign faults at 55% success but fail on gain faults, revealing fundamental limitations of traditional approaches without online parameter estimation.
• →Constant additive bias faults remain unsolved without disturbance observers since integral-free control laws cannot null constant disturbances by definition.
• →Rigorous benchmarking with disjoint train-test splits across inertia, gain, and fault patterns reveals that simulation success claims often overstate real-world robustness.
• →The released benchmark with shared evaluation metrics enables standardized comparison across classical, adaptive, and learned fault-tolerant control methods.