Episodic Transformer for Vision-and-Language Navigation

TL;DR

This paper introduces Episodic Transformer, a multimodal model that encodes language and full episode history to improve vision-and-language navigation in dynamic environments, achieving state-of-the-art results.

Contribution

The paper presents a novel Episodic Transformer architecture that effectively encodes episode history and uses synthetic instructions to enhance navigation performance.

Findings

Encoding episode history with a transformer is crucial for compositional tasks.

Pretraining and synthetic instructions significantly boost navigation success.

Achieves new state-of-the-art on ALFRED benchmark with 38.4% and 8.5% success rates.

Abstract

Interaction and navigation defined by natural language instructions in dynamic environments pose significant challenges for neural agents. This paper focuses on addressing two challenges: handling long sequence of subtasks, and understanding complex human instructions. We propose Episodic Transformer (E.T.), a multimodal transformer that encodes language inputs and the full episode history of visual observations and actions. To improve training, we leverage synthetic instructions as an intermediate representation that decouples understanding the visual appearance of an environment from the variations of natural language instructions. We demonstrate that encoding the history with a transformer is critical to solve compositional tasks, and that pretraining and joint training with synthetic instructions further improve the performance. Our approach sets a new state of the art on the challenging ALFRED benchmark, achieving 38.4% and 8.5% task success rates on seen and unseen test splits.