SpaAct: Spatially-Activated Transition Learning with Curriculum Adaptation for Vision-Language Navigation
SpaAct introduces a training framework for Vision-and-Language Navigation by activating dynamic spatial awareness in Vision-Language Models. It uses two tasks—Action Retrospection and Future Frame Selection—to teach the model backward reasoning about actions and forward prediction of future observations. A progressive curriculum learning strategy called TriPA stabilizes training, leading to state-of-the-art navigation performance.