RS-Claw: Active Tool Exploration via Skill Trees for Remote Sensing Agents
A new paper introduces RS-Claw, a framework that redefines tool selection for remote sensing agents by enabling active exploration within the tool space. Existing passive paradigms, such as full tool registration (Flat) and retrieval-augmented generation (RAG), struggle with balancing context load and toolset completeness in massive, multi-source heterogeneous RS tool ecosystems. Full registration causes context deficits in long-horizon tasks, while RAG may omit critical tools. RS-Claw uses hierarchical skill trees to allow agents to dynamically explore and select tools, improving task reasoning. The paper is published on arXiv under ID 2605.13391.
Key facts
- RS-Claw uses hierarchical skill trees for active tool exploration.
- Existing RS agents use passive selection paradigms: Flat or RAG.
- Full tool registration leads to context space deficits in long-horizon tasks.
- RAG retrieval may omit critical tools in essential steps.
- The paper argues agents should act as active explorers within the tool space.
- The framework targets massive, multi-source heterogeneous RS tool ecosystems.
- The paper is available on arXiv with ID 2605.13391.
- The approach aims to balance context load and toolset completeness.
Entities
Institutions
- arXiv