Combining cultures, from code to canvas: an interview with Ken Goldberg (aihub.org)

In a recent interview, renowned roboticist and artist Ken Goldberg discussed the significant cultural clash within robotics, emphasizing the divide between traditional model-based engineers and those favoring modern learning methods. Goldberg highlighted how the two communities have distinct approaches—specialists building from the ground up versus generalists who assume broad applicability from massive datasets. He advocates for bridging these gaps to optimize the development and deployment of robotics in real-world environments, which is critical as industry demands faster and more reliable systems for revenue generation. Goldberg introduced an innovative "graph-as-policy" concept at ICRA, proposing a hierarchical structure where specialized agents manage individual tasks within a broader framework. This method utilizes large language models (LLMs) to effectively write code for robots, combining the interpretability of traditional methods with the learning capability of modern algorithms. He noted that projects at his lab have demonstrated varying performance among different LLMs, attributing Gemini's superiority to its focus on 3D perception—a crucial element for robotics. As industrial robots generate valuable real-world data, Goldberg suggests that such innovations could lead to enhanced training models, ultimately driving the next wave of robotic capabilities and applications.