Carolina Parada on Gemini Robotics, Robot Data and Physical AGI

General-purpose robotics requires a unified AI layer capable of translating pixels into motor control, moving beyond traditional hard-coded systems. Carolina Parada, VP of Robotics at Google DeepMind, emphasizes that achieving human-level performance involves "embodied reasoning," where models orchestrate complex, multi-step tasks by bridging digital intelligence with physical action. Rather than waiting for full artificial general intelligence, current efforts focus on deploying agent systems—such as those integrated with Boston Dynamics’ Spot—to provide immediate value in structured environments. Training these models relies on diverse, large-scale data, including video and simulation, to ensure robots can handle unstructured, real-world scenarios. By treating robotics as a collaborative evolution of hardware and software, the industry is shifting toward a scalable, foundation-model approach that enables robots to reason about their environment and adapt to novel tasks.