Goosebumps: A Paradigm Shift Is Occurring in Robotics (twitter.com)

A transformative approach named Agentic Robotics (AR) is emerging in the robotics field, shifting the focus from traditional programming and data-heavy learning to multi-agent AI systems that autonomously create, test, and improve robot programs offline. Developed by a collaborative team from UC Berkeley and NVIDIA, AR leverages advanced large language models (LLMs) and visual-language action systems (VLAs) to generate modular robot skills, enabling fast and reliable robot control without extensive pre-existing data. This method combines the interpretability of model-based engineering with the adaptability of model-free approaches, potentially revolutionizing industrial robotic applications. A key innovation within AR is the introduction of Graph-as-Policy (GaP), which organizes control tasks into a graph of modular components that can be concurrently refined by multiple coding agents. This iterative process allows for rapid development and testing, enhancing success rates and efficiency in real-world tasks. However, challenges such as the accuracy of physics simulations and the need for human oversight in fine-tuning remain. The recent performance of the GPT-6 Astra model in creating accurate physics simulations underscores the potential of AR to transform robotic capabilities and reduce reliance on human engineering, indicative of a significant paradigm shift in the field.