Planning with the Views | ViewAgent: View Planning with Multi-turn VLM Agents

ViewAgent: VLM agents that reason and plan over camera views in 3D. ViewSuite, our 6-DoF benchmark for spatial reasoning over views on real ScanNet scenes, shows frontier VLMs track local views but collapse at multi-turn planning; on-policy view graph distillation lifts a 7B VLM agent from 2.5% to…

关键词：ViewAgent, ViewSuite, view planning, VLM agent, view agent, vision-language models, VLM, agentic reinforcement learning, reinforcement learning, on-policy distillation, view graph distillation, self-…