Event
Fudan and Nanyang Technological University propose SimpleVLN navigation method
2 reports · 2 sources
A team from Fudan University, Nanyang Technological University Singapore, and others proposed the SimpleVLN vision-language navigation method: without modifying the core structure of the multimodal large model and without adding dedicated navigation modules, navigation is accomplished solely through observation filtering, action-chunk decision-making, and three-stage training with SFT, DAgger, and reinforcement learning. Each model call serves as an independent decision, taking as input the navigation instruction, the current frame, and filtered historical observations, without accumulating past dialogue and actions. According to the team, after combining R2R, RxR, and ScaleVLN demonstration data, DAgger raised the model's success rate on R2R-CE from 57.9% to 65.4%, and on RxR-CE from 55.3% to 63.4%. The version based on Qwen3-VL 4B takes 293 milliseconds per action-chunk decision on an NVIDIA A100, with latency similar to the streaming method StreamVLN. Beyond simulation evaluation, the team also deployed SimpleVLN on a quadruped robot dog.
Report timeline
- Fudan and NTU propose SimpleVLN: native multimodal large model sets new SOTA in vision-language navigation
A team from Fudan University and Nanyang Technological University, among others, proposed SimpleVLN, which performs vision-language navigation without changing the core structure of a multimodal large model or adding dedicated navigation modules, relying only on observation filtering, action-chunk decision-making, and three-stage training with SFT, DAgger, and reinforcement learning.
CVer · 微信公众号 ·
- Fudan and NTU propose SimpleVLN, using a native multimodal large model to set a new SOTA in vision-language navigation
A team from Fudan University and Nanyang Technological University, Singapore, proposed SimpleVLN, which performs vision-language navigation without changing the core structure of the multimodal large model or adding dedicated navigation modules, relying only on historical frame selection, action-chunk decision-making, and three-stage training with SFT, DAgger, and reinforcement learning.
机器之心 · 微信公众号 ·