Event

SimpleVLN: A VLN framework that preserves the native architecture, leading in success rate

1 reports · 1 sources

Lingyi Hong et al. propose SimpleVLN, a vision-and-language navigation (VLN) framework that adds no task-specific architecture and preserves the native architecture of multimodal large language models (MLLMs). It achieves success rates of 76.7% and 80.1% on the R2R-CE and RxR-CE splits, respectively, which the authors claim is significantly better than existing methods. The framework uses single-turn decision inputs together with a fixed-budget global-local memory to exploit navigation history, avoiding unbounded context growth; for training it adopts a progressive imitation learning-DAgger-reinforcement learning pipeline, combining expert demonstrations, policy-dependent DAgger experience, and a scalable synthetic pipeline to generate large-scale RL tasks. The authors state that it has been instantiated on multiple MLLM backbones and model scales, validating its generality.