VISTA: Scale-Aware Visual Navigation via Action History Conditioning
Researchers have introduced VISTA, a novel approach to visual navigation that addresses the vulnerability of normalized actions in Vision Navigation Foundation Models (VNMs). By conditioning the model on normalized action histories, VISTA provides explicit context on the relationship between predictions and physical displacement, mitigating performance degradation and collision risks. The model also integrates a DINOv3 encoder to better handle visually repetitive environments by capturing spatial and geometric dimensions. VISTA demonstrates robust generalization, achieving 100% goal prediction accuracy in zero-shot real-world deployments across outdoor, forest, and office settings, with an average of 95% checkpoints crossed. AI
IMPACT Enhances robot navigation robustness by conditioning on action history, improving generalization in diverse environments.