Failure-avoidance bayesian optimization with failure-boundary search for automatic exploration of digging control parameters

Bayesian optimization is an effective framework for identifying optimal control parameters of excavation machinery from a limited number of costly trials. While it can improve excavation performance by increasing soil yield and reducing time or fuel consumption, unsafe or off-target parameter combinations must be avoided in real operations. Such cases—e.g., cylinder overload, excessive tire slip, or unstable bucket–soil interaction—are regarded as “failures” in this study. To address this issue, we propose a two-stage safe optimization method designed for automatic excavation control. In the first stage, we explore the success / failure boundary in a simulation environment, where failures can be evaluated without risking damage to vehicles, and estimate the safety probability distribution across the parameter space. In the second stage, this prior safety information is incorporated into a safety-weighted acquisition function for Bayesian optimization, introducing safety as a soft constraint and enabling risk-aware exploration of the parameter space. By separating offline boundary learning and online optimization, the proposed approach aims to reduce failure trials. The effectiveness of the method is evaluated through comparisons with existing safety-aware optimization approaches on known objective functions, as well as through a physics-based excavation simulation.