Robot navigation from human demonstration: learning control behaviors with environment feature maps

TL;DR

This paper introduces a framework enabling unmanned ground vehicles to learn navigation behaviors from minimal human demonstrations using environment feature maps, improving adaptability in dynamic, unstructured settings.

Contribution

The paper presents a novel visual perception and inverse optimal control framework that learns traversal costs from few demonstrations and allows real-time human intervention.

Findings

Requires few demonstration trajectories to learn effective feature costs

Enables real-time human intervention for behavior correction and adaptation

Demonstrates reliable navigation behavior in real-world environments

Abstract

When working alongside human collaborators in dynamic and unstructured environments, such as disaster recovery or military operation, fast field adaptation is necessary for an unmanned ground vehicle (UGV) to perform its duties or learn novel tasks. In these scenarios, personnel and equipment are constrained, making training with minimal human supervision a desirable learning attribute. We address the problem of making UGVs more reliable and adaptable teammates with a novel framework that uses visual perception and inverse optimal control to learn traversal costs for environment features. Through extensive evaluation in a real-world environment, we show that our framework requires few human demonstrated trajectory exemplars to learn feature costs that reliably encode several different traversal behaviors. Additionally, we present an on-line version of the framework that allows a human teammate to intervene during live operation to correct deteriorated behavior or to adapt behavior to dynamic changes in complex and unstructured environments.