Title:Normalizing Flows are Capable Visuomotor Policy Learning Models

Abstract:The field of general purpose robotics has recently embraced powerful probabilistic models, such as diffusion models, to model and learn complex behaviors. However, these models often come with significant trade-offs, namely high computational costs for inference and a fundamental inability to quantify output uncertainty. We argue that a model's trustworthiness, a critical factor for reliable, general-purpose robotics, is inherently linked to its ability to provide confidence measures.
In this work, we introduce Normalizing Flows Policy, a novel visuomotor policy learning model based on Normalizing Flows. We show that Normalizing Flows are a natural and powerful alternative to diffusion models, providing both a statistically sound measure of confidence and a highly efficient inference process. Through comprehensive experiments across four distinct simulated robotic tasks, we demonstrate that Normalizing Flows Policy achieves performance comparable to, and often surpassing, Diffusion Policy, and it does so not only with improved sample efficiency but also with up to 30 times faster inference. Additionally, our ablation study validates several key architectural and training techniques that enable Normalizing Flows to perform well in this domain.