This repository provides the starter code and infrastructure for the 2023 Habitat Navigation Challenge, focusing on ObjectNav and ImageNav tasks. It's designed for researchers and developers in embodied AI aiming to build agents that can navigate unseen environments to find specific objects or match goal images, leveraging realistic robot configurations and continuous action spaces.

How It Works

The challenge utilizes the HM3D-Semantics v0.2 dataset and simulates navigation for a HelloRobot Stretch robot. Agents interact with the environment using continuous actions (linear/angular velocity, camera pitch) or abstract waypoint commands. Performance is evaluated using Success and Success weighted by Path Length (SPL) metrics, assessing both task completion and navigation efficiency.

Quick Start & Requirements

• Install: Clone the repository (git clone https://github.com/facebookresearch/habitat-challenge.git) and install dependencies via Docker.
• Prerequisites: Linux OS, NVIDIA Docker v2, CUDA.
• Dataset: Download HM3D-Semantics v0.2 and place it in habitat-challenge-data/data/scene_datasets/hm3d_v0.2.
• Local Testing: Use scripts/test_local_objectnav.sh or scripts/test_local_imagenav.sh after building a Docker image.
• Resources: Requires significant disk space for datasets and GPU resources for training/evaluation.
• Tutorials: Video and Colab tutorials are available.

Highlighted Details

• Two tasks: ObjectNav (find object category) and ImageNav (find specific object instance via goal image).
• Continuous action space for sim-to-real transfer, with optional waypoint controllers.
• Evaluation metrics: Success and SPL, with oracle visibility for success determination.
• Baseline DD-PPO training code and pre-trained weights are provided.

Maintenance & Community

The project is maintained by Facebook AI Research (FAIR). Questions and issues can be raised via GitHub issues.

Licensing & Compatibility

The code is released under a permissive license, allowing for commercial use and integration with closed-source projects.

Limitations & Caveats

The provided starter code and Docker setup are Linux-specific. Participants must manage dataset downloads and Docker image updates. Overfitting to the test set is discouraged due to limited submissions for leaderboard and challenge phases.