Survey on safe robot control via learning

TL;DR

This survey addresses safe robot learning by mapping the spectrum between classical control and learning-based methods, and by identifying safety guarantees such as stability and reachability that must be preserved in data-driven controllers. It covers model-free and model-based reinforcement learning, robustness challenges like sim-to-real transfer, and augmentations of classical control with neural models, including Neural Lander, Neural Lyapunov with dReal, and SINDY. The work highlights how temporal logics and differentiable specifications can express safety properties, and it discusses embedded system considerations—latency, unpredictability, and resource constraints—that critically impact safety in real hardware. Overall, the article emphasizes the need for verifiable guarantees, scalable verification tools, and cross-stack integration to enable practical, safe deployment of learning-enabled robotics. The findings point to promising directions in online learning with safety guarantees, Lyapunov-certified NN controllers, and model-based RL, while acknowledging computational and generalization challenges that require further research.

Abstract

Control systems are critical to modern technological infrastructure, spanning industries from aerospace to healthcare. This survey explores the landscape of safe robot learning, investigating methods that balance high-performance control with rigorous safety constraints. By examining classical control techniques, learning-based approaches, and embedded system design, the research seeks to understand how robotic systems can be developed to prevent hazardous states while maintaining optimal performance across complex operational environments.