ImitationFlow: Learning Deep Stable Stochastic Dynamic Systems by\n Normalizing Flows

2020/10/25 by Julen Urain, Urain, Julen, Michelle Ginesi +5 · 4 citations
Physics and Astronomy · Computer Science · #Model Reduction and Neural Networks #Reinforcement Learning in Robotics #Generative Adversarial Networks and Image Synthesis

paper · pdf · doi:10.48550/arxiv.2010.13129

Abstract

We introduce ImitationFlow, a novel Deep generative model that allows\nlearning complex globally stable, stochastic, nonlinear dynamics. Our approach\nextends the Normalizing Flows framework to learn stable Stochastic Differential\nEquations. We prove the Lyapunov stability for a class of Stochastic\nDifferential Equations and we propose a learning algorithm to learn them from a\nset of demonstrated trajectories. Our model extends the set of stable dynamical\nsystems that can be represented by state-of-the-art approaches, eliminates the\nGaussian assumption on the demonstrations, and outperforms the previous\nalgorithms in terms of representation accuracy. We show the effectiveness of\nour method with both standard datasets and a real robot experiment.\n

Cited by

• LILAD: Learning In-context Lyapunov-stable Adaptive Dynamics Models
• An Alignment-Based Approach to Learning Motions from Demonstrations
• A Robust Task-Level Control Architecture for Learned Dynamical Systems
• KoopMotion: Learning Almost Divergence Free Koopman Flow Fields for Motion Planning

Related

• Learning Lyapunov-Stable Polynomial Dynamical Systems through Imitation
• Learning Stable Deep Dynamics Models
• Learning Deep Dynamical Systems using Stable Neural ODEs
• Stable Autonomous Flow Matching
• How to Learn and Generalize From Three Minutes of Data: Physics-Constrained and Uncertainty-Aware Neural Stochastic Differential Equations