A Flexible Imitation Learning System Based on the Lyapunov Energy Function of the Neural Networks

Endowing robots with the capability to learn skills from human demonstrations is cutting-edge research. Traditional data-driven neural networks are typically employed to learn functions that represent dynamical systems for model skills. Despite the effectiveness of these networks, it has been difficult to guarantee that the learned function has a unique minimum. Although some algorithms attempt to improve it by imposing strong constraints, they cannot capture demonstration prefer-ences well due to structural limitations. In this paper, the specially designed residual network is introduced into the Lyapunov function and dynamical systems, ensuring that the system has a unique minimum during data-driven learning. Due to the residual network's ability to learn identity mapping, we have incorporated feature guidance in the network's input, leading to a significant improvement in the learning performance of the neural network. Experimental results demonstrate that our system outperforms traditional algorithms on the LASA handwritten dataset.