Vision-and-language navigation tasks often push researchers toward specialized model designs. In a new arXiv preprint, the authors of SimpleVLN argue that this may be unnecessary. Their framework keeps the underlying multimodal large language model (MLLM) architecture intact, instead focusing on how navigation is formulated, how data is built, and how optimization is staged.

The formulation uses single-turn decision inputs paired with a fixed-budget global-local memory, so the model draws on navigation history without letting context grow without bound. Action chunks let the model output several primitive actions in one call. On the data side, the authors combine expert demonstrations with policy-dependent DAgger experience and build a scalable synthesis pipeline to generate large-scale tasks for reinforcement learning.

Optimization proceeds through a progressive imitation-learning-to-DAgger-to-RL pipeline, moving from action imitation to task-level outcome optimization. The authors instantiate SimpleVLN across multiple MLLM backbones and model sizes, reporting 76.7% and 80.1% success rates on the R2R-CE and RxR-CE splits. Their central claim is that native MLLMs can be effectively unlocked for navigation without introducing task-specific architectural components.