Abstract: Accurate landing detection is crucial for humanoid robots performing high dynamic motions. Unlike common methods that rely on redundant force-torque sensors and low-precision observers to estimate landing states, this paper proposes a novel landing detection method characterized by high precision and low noise, synthesizing a learning-based Improved Momentum Observer (IMO-Net) for the ankles' external torque estimation with a Gated Recurrent Unit (GRU)-based network for state judgment. Since the movement and external torque of the ankle undergo drastic changes during high dynamic motions, achieving accurate and real-time estimation presents a challenge. To address this problem, IMO-Net employs a new Improved Momentum Observer(IMO), which does not depend on acceleration data derived from second-order differentials or friction model, and significantly reduces noise effects from sensors data and robot foot wobble. Fur-thermore, an Elman network is utilized to accurately calculate the ankle output torque(IMO input), significantly reducing the estimation error. Finally, leveraging IMO-Net and extensive experimental data, we developed and optimized a GRU-based landing detection network through comprehensive ablation experiments. This refined network reliably determines the robot's landing states in real-time. The effectiveness of our methods has been validated through experiments

Key words: External torque estimation')">External torque estimation, Landing detection, Improved momentum observer, GRU-based