Multi-fidelity Optimal Trajectory Generation: Optimal Experiment Design for Robot Learning

Abstract

Data-driven methods have significantly advanced robot learning, yet their direct application to real-world robots remains challenging, particularly under extreme conditions. This challenge is especially pronounced for highly maneuverable vehicles like quadrotor aircraft, which often operate in scenarios requiring rapid maneuvering, such as racing, defense systems, or safety-critical obstacle avoidance. In such extreme conditions, real-world constraints like control delays, state estimation errors, and battery voltage fluctuations often compromise trajectory reliability, even when conforming to ideal dynamics. However, the typical data-driven methods are usually developed in simulated environments. Consequently, the transition to real-world dynamics requires extensive fine-tuning, which can be risky, as perfect training in simulations does not guarantee safe transitions to real-world dynamics. This thesis employs methods from optimal experiment design to address these challenges. By quantifying uncertainty and maximizing information gain, the approach aims to safely and efficiently design the real-world experiments required for accurate constraint modeling. In the first chapter, we present a multi-fidelity Bayesian optimization method that searches for time-optimal speed profiles for quadrotor aircraft, effectively balancing numerical simulations with real-world flight experiments. The second chapter extends the optimal experiment design method to a high-dimensional online planning problem through integration with reinforcement learning. The proposed algorithms, trained and validated through real-world flight experiments, significantly outperform baseline methods in trajectory time and computational efficiency. Additionally, these algorithms have been adapted to various planning problems, including fixed-wing aircraft planning, cooperative multi-drone systems, and energy-efficient trajectory generation.