EmbodiedTool: Enabling Extensible Embodied Capabilities with Tools

Compared with end-to-end embodied models that jointly learn perception, reasoning, planning, and control, augmenting foundation models with external tools offers a flexible alternative, allowing specialized capabilities to be independently optimized, reused, and invoked on demand. Recent Code-as-Policy methods have demonstrated the promise of this paradigm, yet progress remains limited by the lack of a unified interface for heterogeneous tools, a comprehensive reusable tool base, and systematic evaluation of models' tool-use capabilities. To bridge these gaps, we present EmbodiedTool, which provides a unified agent-facing interface, integrates 112 validated tools spanning perception, cognition, reasoning, and execution, and introduces EmbodiedToolBench to evaluate tool-necessity recognition, tool selection, tool execution, and tool-chain composition. Experiments across simulation and real-world platforms show that tool augmentation can substantially improve embodied performance, achieving average gains of 31% on EB-ALFRED and 36% on EB-Navigation, while revealing a clear capability-dependent boundary: gains are substantial for perception and cognition but remain limited for execution-oriented capabilities. Meanwhile, current models still struggle with when, which, and how to use embodied tools, highlighting embodied tool competence as a key challenge for tool-augmented embodied intelligence.