Beyond the Nav-Graph: Vision-and-Language Navigation in Continuous Environments

We develop a language-guided navigation task set in a continuous 3D\nenvironment where agents must execute low-level actions to follow natural\nlanguage navigation directions. By being situated in continuous environments,\nthis setting lifts a number of assumptions implicit in prior work that\nrepresents environments as a sparse graph of panoramas with edges corresponding\nto navigability. Specifically, our setting drops the presumptions of known\nenvironment topologies, short-range oracle navigation, and perfect agent\nlocalization. To contextualize this new task, we develop models that mirror\nmany of the advances made in prior settings as well as single-modality\nbaselines. While some of these techniques transfer, we find significantly lower\nabsolute performance in the continuous setting -- suggesting that performance\nin prior `navigation-graph' settings may be inflated by the strong implicit\nassumptions.\n