Robot-Use Agents: Why General-Purpose Models May Win in Robotics

General-purpose LLMs are transforming robotics by acting as "robot-use agents" that generate code to control physical hardware. This shift moves away from rigid, fine-tuned Vision-Language-Action models toward flexible systems that leverage in-context learning and tool use to execute complex tasks. The "Platonic Representation Hypothesis" suggests that as models scale, they develop consistent internal mappings of the physical world, allowing them to generalize across domains like coding, computer use, and robotics. While current approaches face latency bottlenecks, integrating computer-use data and refining software harnesses for skill consolidation promise to accelerate development. Experts anticipate that general-purpose robots capable of performing human-like tasks will emerge within two years, necessitating new methods for distilling experiences into efficient, high-throughput policies that mimic biological learning cycles.