Language-Conditioned World Modeling for Visual Navigation
Language-Conditioned Visual Navigation (LCVN) addresses generating navigation paths based only on visual input and natural language instructions, without goal images. The research proposes two frameworks, LCVN-WM and LCVN-AC, which use world models to predict future states conditioned on language. Results show that unified architectures generalize better to new environments, with landmark-based instructions proving most effective for guiding agents.