ManifoldTechLtd/Odin-Nav-Stack

What it solves

This project provides a comprehensive navigation and intelligence stack for ground robots (specifically the Unitree Go2) using the Odin1 spatial sensing module. It addresses the challenge of enabling robots to navigate complex environments autonomously while understanding their surroundings through semantic object detection and natural language interaction.

How it works

The stack integrates several components into a modular ROS1-based architecture:

• Sensing & Mapping: Uses the Odin1 module for 3D mapping and persistent relocalization.
• Path Planning: Employs a high-performance neural planner (NeuPAN) for local obstacle avoidance and global path planning, featuring a stuck-escape strategy and dynamic obstacle memory.
• Semantic Understanding: Integrates YOLOv5 for object detection and a Vision-Language Model (VLM) via LLaMA.cpp to generate natural language descriptions of scenes.
• Navigation Control: Supports Vision-Language Navigation (VLN), allowing the robot to move to specific objects based on spoken or typed commands.

Who it’s for

Robotics developers and researchers working with the Unitree Go2 or similar ground platforms who need a production-ready navigation stack that combines traditional SLAM with modern AI-driven scene understanding.

Highlights

• Dynamic Obstacle Avoidance: Reactive local planners designed for safe motion in complicated environments.
• Natural Language Navigation: Ability to detect, localize, and navigate to objects using commands like "Go to the left of the chair."
• Multimodal AI Integration: Uses VLMs to provide contextual scene descriptions.
• Neural Planning: Implements the NeuPAN end-to-end neural planner for high-performance local navigation.
• Modular Deployment: Supports Docker for quick deployment on NVIDIA Jetson hardware.