started · updated
UC Berkeley researchers develop agentic robotics and local AI models
Researchers at UC Berkeley have introduced a new approach called agentic robotics designed to accelerate how robots learn and perform tasks. This method utilizes teams of AI agents to generate control programs, test them within physics-based simulations, and refine code before physical deployment. By combining agentic AI with modular control systems, the approach aims to reduce the heavy reliance on manual engineering and the massive datasets typically required for imitation learning.
In a separate development, researchers from UC Berkeley and the University of Colorado Boulder are advocating for the use of smaller, open-weight AI models hosted on local campus servers. This strategy allows scientific researchers to maintain control over sensitive data and ensures the reproducibility of results, which can be compromised if commercial chatbot providers update or retire their models. The authors suggest that smaller models are often sufficient for specific research tasks, such as data extraction or code assistance, potentially reducing the need for massive computing investments.