ZeroVLN navigates without SFT of the VLM brain using tool-calling and closed-loop RL
ZYT-AI · hf · 2026-11-20
ZeroVLN, from Beihang University, Zhuoyu Technology and Zhejiang University, is a slow-fast vision-language navigation framework that avoids navigation-domain SFT of the VLM brain. It uses a semantic tool-calling interface to convey movement decisions in natural language to an action expert, then jointly optimizes the brain's decisions and the expert's execution via closed-loop RL. It substantially outperforms SFT-based hierarchical VLN methods on VLN-CE while preserving the VLM's general capabilities, and is the first to jointly optimize slow-fast policies with agentic RL in VLN-CE.
More from Embodied
- NVIDIA's SOMA Retargeter adapts human motion to five humanoid robots, from G1 to AgiBot — jonstephens85 · 2026-11-20
- AFFECT from KAIST makes world models dynamics-faithful via action effect matching — John Won · 2026-11-20
- Mind2Act World: a 180-task benchmark exposing robots' mind-action composition failures — Mind2Act-World · 2026-11-20
- Robot Paper Club recordings go live with four post-training talks; World Models edition next — DominiqueCAPaul · 2026-11-20
- Chinese student builds wearable tactical backpack robot, demo goes viral — TinfoilTricorn · 2026-11-20
- Huawei Cloud open-sources Ascend-native embodied world model CloudRobo-WM1.2, topping two benchmarks — 量子位 · 2026-11-20