Recent breakthroughs in artificial intelligence are driving a transformative new approach to robot programming called agentic robotics, announced by UC Berkeley professor Ken Goldberg and collaborators from NVIDIA. This method uses teams of AI agents that autonomously compose robot-control software, test it in simulations, and iteratively improve performance without requiring vast data collection or laborious manual coding.
Agentic robotics bridges the two traditional paradigms in robotics: model-based methods, which rely on handcrafted software and extensive engineering effort, and model-free learning, which depends on large datasets and time-consuming training. Goldberg calls it "a fundamental shift" as it eliminates the need for accumulating immense real-world data typically necessary for effective robot behaviors.
Core to this approach is the Graph-as-Policy framework, which decomposes complex robotics tasks into modular skills—such as object recognition, grasp planning, and motion execution. AI agents then synthesize and link these skill blocks into full programs, automatically testing and refining them using NVIDIA’s Isaac simulation environment before deployment in physical robots.
The team demonstrated these capabilities by leveraging OpenAI’s Astra model, enabling AI to create realistic physics simulations from brief real-world video, and subsequently generate refined robot-control code for tasks like wiping ink from metal with a sponge. Simulated failures inform iterative improvements, highlighting the system's autonomous self-correction capabilities.
Industrial implementations are already benefiting from agentic robotics. Berkeley-based Ambi Robotics adopted the approach to analyze and enhance parcel sorting robots, resulting in a throughput gain exceeding 15,000 packages annually per robot. These improvements are actively deployed in nearly a third of the company's U.S. robot fleet, underscoring practical scalability.
Despite promising industrial results, agents still face limitations in generalization and the nuanced complexities of open environments, suggesting home assistant robots performing intricate duties remain a future challenge. The focus currently stays on repetitive, scalable tasks common in factories and logistics.
Goldberg emphasizes the synergy between classical robotics control theory and advanced agent-based AI coding, branding agentic robotics as a rapid, practical methodology poised to reshape the engineering of industrial robot systems. He likens its significance to landmark milestones like the Internet’s public launch and ChatGPT’s debut.
Open-source releases of their frameworks and benchmarks aim to accelerate community adoption and development, positioning agentic robotics as a foundational step towards more autonomous, adaptable robotic systems in the near term.
Source & further reading
Recent advances in AI are creating a ‘paradigm shift’ for robotics ↗Source published 2026-11-09. This report was prepared with AI assistance from the linked primary-source material. Claims and limitations should be checked against the original announcement or research.