Recent advances in AI are creating a ‘paradigm shift’ for robotics
AI-driven agentic robotics enables robots to self-improve control programs through simulation and real-world testing, reducing reliance on massive datasets and hand-coding.
Professor Ken Goldberg of UC Berkeley describes recent AI advancements as a transformative shift for robotics, citing new models that allow teams of AI agents to generate, test, and refine robot-control programs autonomously. This approach, termed 'agentic robotics,' addresses longstanding challenges in the field by minimizing the need for extensive data collection and manual programming, which traditionally demand significant physical and technical resources.
Goldberg and collaborators, including NVIDIA’s Spencer Huang, presented the method at the Conference on Robot Learning and released open-source code built on NVIDIA’s Isaac simulation framework. The team also introduced an industry-standard benchmark for evaluating robot performance on industrial tasks. The goal is to accelerate robot programming while maintaining reliability, making development faster and more adaptable for real-world applications.
The new framework, Graph-as-Policy, segments robot tasks into modular skills that AI agents assemble into control programs. Using NVIDIA’s Isaac simulation libraries, these agents autonomously rehearse and refine programs before deployment. A demonstration with OpenAI’s Astra model showed the AI creating a physics simulation from a video of a robot wiping ink off a metal bar, generating and improving a control program that worked on a real robot.
Industrial robotics company Ambi Robotics applied agentic robotics to analyze parcel sorting data, resulting in a 15,000-package annual increase in sorting capacity per robot. The improvement is now deployed across 30% of the company’s U.S. fleet, demonstrating practical benefits in logistics and warehouse operations.