NavBEST: Behavior-Enhanced Strategy With Spatio-Temporal Perception for Mapless Navigation in Dynamic Environments

Le résumé fourni par la source

Autonomous navigation in narrow and dynamic indoor environments remains highly challenging for learning-based methods. Existing approaches still face issues such as limited observation and convergence complexity. In light of this, we propose NavBEST, which consists of a spatio-temporal perception (STP) module and a behavior-enhanced strategy (BES) module. The STP guides predictive visual representations to align with goal states. This enables the navigation policy to anticipate potential collisions and effectively prune unsafe action choices, without relying on precise instance-level segmentation. In BES, imitation learning (IL) is initially trained for obstacle avoidance in static scenarios, offering alternative action to guide reinforcement learning (RL) bootstrapping in dynamic environments. The BES evaluates Q-values for both actions, maps them into probabilities, and samples the optimal action for RL updates, enhancing training efficiency by introducing more positive samples. As a result, STP yields goal-oriented latent representations of current and future observations using only RGB input, while BES adaptively balances action selection between IL and RL throughout training. Extensive validations in both simulation and real-world demonstrate our method's superiority over state-of-the-art approaches. Videos of our experiments are available athttps://youtu.be/ulET2BhlUSQ.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.