Learning Globally Stable Neural Imitation Policies via Reshaped Energy Functions

Energy function-based DSs hold tremendous potential for developing point-to-point motion learning models with stability and accuracy. Nevertheless, a commonly encountered issue is that most methods struggle to strike a balance between stability and accuracy. This article proposes a novel adaptive modulation scheme that aims to learn a DS based on neural energy functions to alleviate the constraints of traditional methods. In this method, we decouple and modulate the radius and angle of the motion position in a polar-like space through coordinate transformation to ensure global stability and learning accuracy for point-to-point motion learning and generalization. Both simulations using a public benchmark and real-world evaluations of ultrasound robot scanning demonstrate that the proposed method exhibits superior generalization and higher accuracy for motions compared to existing motion learning methods.