Robust autonomous navigation in complex, dynamic indoor environments remains a central challenge in robotics, requiring agents to make adaptive decisions in real time under partial observability and uncertain obstacle motion. This paper presents DreamerNav, a robot-agnostic navigation framework that extends DreamerV3, a state-of-the-art world model–based reinforcement learning algorithm, with multimodal spatial perception, hybrid global–local planning, and curriculum-based training. By formulating navigation as a Partially Observable Markov Decision Process (POMDP), the system enables agents to integrate egocentric depth images with a structured local occupancy map encoding dynamic obstacle positions, historical trajectories, points of interest, and a global A* path. A Recurrent State-Space Model (RSSM) learns stochastic and deterministic latent dynamics, supporting long-horizon prediction and collision-free path planning in cluttered, dynamic scenes. Training is carried out in high-fidelity, photorealistic simulation using NVIDIA Isaac Sim, gradually increasing task complexity to improve learning stability, sample efficiency, and generalization. We benchmark against NoMaD, ViNT, and A*, showing superior success rates and adaptability in dynamic environments. Real-world proof-of-concept trials on two quadrupedal robots without retraining further validate the framework’s robustness and quadruped robot platform independence.
Shanks et al. (Fri,) studied this question.