Title: Leveraging Reinforcement Learning for Optimal Experiment Design in Mechatronic Parameter Identification

Abstract: Generating informative excitation signals is essential for the precise system identification of mechatronic devices. However, traditional system identification (SI) methods often rely on manual signal design and specialized expert knowledge to ensure compliance with hardware safety limits, which restricts their broader applicability. To address this, we introduce a reinforcement learning (RL) agent capable of autonomously deriving optimal excitation signals for a Quanser Aero 2 testbed. By utilizing reward shaping, the agent strictly adheres to safety constraints without human intervention. Our comprehensive evaluation, conducted across ten independent training seeds, demonstrates that the proposed agent delivers competitive estimation accuracy for all three targeted parameters. It not only surpasses classical baseline methods but also maintains a minimal safety violation rate of just 0.75%.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC