<2-3 sentences summarising what the project addresses and solves, the target audience, and the benefit.> ApexNav provides a reliable and efficient framework for zero-shot object navigation in complex environments. Targeting robotics researchers and engineers, it enables autonomous agents to navigate towards objects without prior specific training, enhancing adaptability and reducing development overhead. The system leverages advanced semantic fusion and adaptive exploration strategies for robust performance.

How It Works

ApexNav employs a novel approach combining Target-centric Semantic Fusion for robust object recognition and localization with an Adaptive Exploration Strategy to efficiently navigate unknown environments. This synergy allows for zero-shot object navigation, enabling agents to locate and reach target objects without prior specific training on their appearance, by dynamically adjusting exploration based on semantic cues and environmental feedback.

Quick Start & Requirements

• Primary Install/Run: Requires ROS Noetic and Python 3.9, with Anaconda/Miniconda recommended for environment management.
• Prerequisites: System dependencies include libarmadillo-dev and libompl-dev. Optional LLM setup involves Ollama with the qwen3:8b model. External code dependencies include yolov7 and GroundingDINO.
• Model Weights: Download mobile_sam.pt, groundingdino_swint_ogc.pth, and yolov7-e6e.pt.
• Environment Setup: Create a Conda environment using apexnav_environment.yaml. PyTorch installation is CUDA-specific (versions for 11.8, 12.1, 12.4 provided). Habitat simulator (v0.3.1) and baselines are required. Specific numpy (1.23.5) and numba (0.60.0) versions are noted.
• Datasets: Requires HM3D and MP3D scene datasets (application/download needed) and ObjectNav datasets (HM3D-v0.1, HM3D-v0.2, MP3D-v1). Links for Habitat datasets are available at https://github.com/facebookresearch/habitat-lab/blob/main/DATASETS.md.
• Execution: Involves ROS compilation (catkin_make), running VLM servers (Grounding DINO, BLIP-2, SAM, YOLOv7), and launching visualization (rviz.launch) or the main algorithm (exploration.launch). Evaluation and keyboard control scripts are also provided.
• Real-world: A real-world deployment example is available, detailed in a separate README.

Highlighted Details

• Published in IEEE Robotics and Automation Letters (RA-L'25).
• Features a real-world deployment example code.
• Supports zero-shot object navigation via Target-centric Semantic Fusion and Adaptive Exploration.
• Integrates with the Habitat simulator for evaluation and keyboard control.

Maintenance & Community

No explicit community channels (Discord, Slack) or roadmap links are provided. Project development is active, with releases and news noted in late 2025.

Licensing & Compatibility

The repository's license is not explicitly stated in the README. This requires clarification for commercial use or integration into proprietary systems.

Limitations & Caveats

Current implementation relies on ROS Noetic; ROS2 support is a future TODO item. Setup involves obtaining and configuring potentially restricted scene datasets (HM3D, MP3D), which may require significant effort and permissions. The project appears research-oriented, with potential for evolving APIs or features.