DreamerNav: learning-based autonomous navigation in dynamic indoor environments using world models.

Area of Science:

• Robotics
• Artificial Intelligence
• Reinforcement Learning

Background:

• Autonomous navigation in dynamic indoor environments is a significant challenge.
• Real-time decision-making under partial observability and uncertain obstacle motion is crucial for robotic agents.

Purpose of the Study:

• To introduce DreamerNav, a robot-agnostic navigation framework.
• To enhance state-of-the-art reinforcement learning algorithms with multimodal perception and planning strategies for improved navigation.

Main Methods:

• Formulating navigation as a Partially Observable Markov Decision Process (POMDP).
• Integrating egocentric depth images with a structured local occupancy map using a Recurrent State-Space Model (RSSM).
• Employing curriculum-based training in high-fidelity simulation (NVIDIA Isaac Sim).

Main Results:

• DreamerNav demonstrated superior success rates and adaptability compared to NoMaD, ViNT, and A* in dynamic environments.
• The framework achieved collision-free path planning in cluttered and dynamic scenes.
• Real-world trials on quadrupedal robots confirmed the framework's robustness and platform independence without retraining.

Conclusions:

• DreamerNav offers a robust and adaptable solution for autonomous navigation in challenging indoor environments.
• The multimodal perception and hybrid planning approach significantly improves navigation performance.
• The robot-agnostic nature and successful real-world validation highlight the framework's practical applicability.