InternRobotics/NavDP

What it solves

NavDP addresses the challenge of transferring navigation capabilities from simulation to the real world (sim-to-real) without requiring real-world robot data. It enables robots to perform real-time path planning and obstacle avoidance across different robot types (cross-embodiment generalization) in various tasks such as exploring without a goal, navigating to a specific point, or navigating to a target image.

How it works

NavDP is an end-to-end mapless navigation model that utilizes a diffusion policy. It employs a highly efficient simulation data generation pipeline and a specific model design guided by privileged information to learn navigation behaviors. The system is designed to be decoupled, where the navigation model runs as a server that receives RGB-D observations and generates preferred trajectories, which are then tracked by an MPC-based controller.

Who it’s for

This project is intended for robotics researchers and developers working on embodied AI, specifically those focused on visual navigation, sim-to-real transfer, and the deployment of diffusion policies for robot motion planning.

Highlights

• Cross-Embodiment Generalization: Works across various robot platforms without needing real-world data.
• Multi-Task Support: Handles no-goal exploration, point-goal navigation, and image-goal navigation.
• High-Fidelity Benchmarking: Includes a platform based on IsaacSim and IsaacLab with realistic physics and scene assets to minimize the sim-to-real gap.
• Decoupled Architecture: Separates the navigation approach from the evaluation process via HTTP requests for easier testing of new methods.