AtlasVA

AtlasVA: Self-Evolving Visual Skill Memory for Teacher-Free VLM Agents

This repository contains the model weights for AtlasVA, as presented in the paper AtlasVA: Self-Evolving Visual Skill Memory for Teacher-Free VLM Agents.

AtlasVA is a teacher-free visual skill memory framework designed for Vision-Language Model (VLM) agents. It organizes memory into three complementary layers: spatial heatmaps, visual exemplars, and symbolic text skills. By evolving danger and affinity atlases directly from trajectory statistics, AtlasVA provides dense, coordinate-aware guidance for reinforcement learning, unifying perception, memory, and optimization without external LLM supervision.

• Project Page: https://wangpan-ustc.github.io/AtlasvaWeb/
• Repository: https://github.com/wangpan-ustc/AtlasVA
• Paper: https://huggingface.co/papers/2605.17933

Model Details

• Base Model: Qwen2.5-VL-3B-Instruct
• Task: Multimodal agentic decision making (Sokoban, FrozenLake, 3D navigation, robotic manipulation).
• Memory Layers: Spatial heatmaps, visual exemplars, and symbolic text skills.

Citation

@article{wang2026atlasva,
  title={AtlasVA: Self-Evolving Visual Skill Memory for Teacher-Free VLM Agents},
  author={Wang, Pan and Hu, Yihao and Liu, Xiujin and Yang, Jingchu and Wang, Hang and Wen, Zhihao},
  journal={arXiv preprint arXiv:2605.17933},
  year={2026}
}