Should robots outsource their brains to tools?

Current embodied AI systems jam perception, reasoning, and control into one model, making them brittle. This work decouples capabilities into independent tools—specialized modules for seeing, thinking, and acting—that a robot calls on as needed. Testing on navigation and household tasks shows 31–36% gains, but reveals the real bottleneck: models are poor at deciding *when* and *which* tools to invoke, not at executing them.