VLN-CE is a vision-and-language navigation task where agents navigate through complex, continuous environments guided by natural language instructions. This repository serves as a starting point for researchers and developers to explore and contribute to this challenging task. We implement baseline agents and provide training methods using the Habitat platform, focusing on the Room-to-Room (R2R) and Room-Across-Room (RxR) datasets. The primary objective is to advance the state-of-the-art in embodied AI by developing agents that can understand and execute complex instructions in dynamic environments.
This project distinguishes itself by supporting both R2R and RxR datasets, offering a comparative evaluation platform for VLN-CE agents. The implementation leverages the Habitat simulator, providing a standardized environment for training and evaluation. The code is designed for modularity, allowing for easy integration of new architectures and training algorithms. It supports multiple datasets and is actively maintained with updates and improvements.
- Core Functionality: Provides implementations for training and evaluating vision-and-language navigation agents in Habitat.
- Supported Platforms: Supports both R2R and RxR datasets.
- Configuration: Uses a configurable YAML file for model and training parameters.
- Performance: Designed for scalable training with support for distributed environments.
- Developer Experience: Offers clear documentation, example usage scripts, and a modular codebase.
The project is actively maintained with ongoing development and improvements. It is based on the Habitat platform, which has a strong community and active development. The repository includes documentation and example usage. Regular updates are provided to address bugs and incorporate new features, with a focused effort on improving performance and usability. The provided example code allows for onboarding
This project benefits researchers and developers interested in vision-and-language navigation. It provides a comprehensive platform for training and evaluating agents in continuous, complex environments, supporting both R2R and RxR datasets. It facilitates research and development in embodied AI by offering a robust and well-documented implementation built on the Habitat simulator.