UC Berkeley and NVIDIA Researchers Use AI Agents to Improve Robot Control Code in Virtual Environments

TechXplore reported on Nov. 9, 2026, that UC Berkeley robotics researcher Ken Goldberg, NVIDIA Director of Robotics Spencer Huang and others proposed “agentic robotics,” in which groups of AI agents develop and test robot control code. The agents try the code in a simulated environment and revise it if it does not work. Goldberg said he plans to present the method along with results from several months of testing.

The researchers say the method aims to address the large amounts of data needed for robot development and the burden of manually tuning behavior. How well code revisions in simulation translate to performance and safety on physical robots remains a key question in assessing the method’s scope; the article did not provide specific results on this point.

The focus of the research is not just on increasing the data used to teach robots how to move, but on incorporating groups of agents into the process of writing, testing and revising control programs. The article describes a development process in which AI agents test code in simulation and revise it when it fails; it does not describe a system for autonomous operation on physical robots or one that eliminates the need for human review. Goldberg’s account also leaves human engineers responsible for interpreting how the programs work, making adjustments and verifying them. The researchers say that large volumes of training data and manual adjustments for individual behaviors are challenges in conventional robot development. Their proposal is an approach to addressing these burdens through code generation and iterative testing in simulation. However, the article excerpt does not identify the robots tested, the simulator, the number of agents or comparative results against conventional methods. It is therefore not possible to determine how much development time or required data might change. In a related article, NVIDIA Presents Robot Research at CoRL on Learning Precision Tasks in Simulation, we covered research in which robots practice manual tasks such as gear alignment and cable connection in simulation. Unlike those methods for learning specific tasks, “agentic robotics” proposes using agents in the development and iteration of control code itself. Both approaches use simulation, but their target processes and measures of success should be considered separately. The implication for the industry is that simulation could become not only a place to test robot behavior but also a development environment for generating and verifying control software. To assess its practicality, it will be necessary to examine the success rate, safety and reproducibility of code revised in simulation when transferred to physical robots, as well as the amount of work engineers need to spend on verification. The article mentions several months of testing but does not provide quantitative results on these points or say whether the work was deployed on physical robots. It is important not to equate a research-stage proposal with automation ready for operational use.