UC Berkeley and NVIDIA release AI agent development tool for robots, report says

The Robot Report reported on November 11, 2026, that UC Berkeley and NVIDIA released an open-source tool that uses AI agents to create, simulate and improve robot programs.

According to The Robot Report, the tool supports development in which AI agents generate robot programs, then test and improve them in simulation. For industrial robot developers, the readability of generated code and the verification process are among the factors that determine whether it can be used in the field.

The technical focus of the report is not only on models that directly produce robot movements, but also on a development process that supports the creation and verification of readable programs. The reported role of agents in testing and improving programs as well as generating them could broaden the unit of evaluation for robot AI from the model alone to software outputs that people can inspect and the development process. However, this is an interpretation of the reported concept and should be distinguished from results showing improved performance on physical robots or reduced maintenance burdens. In a report published the previous day on test results for NVIDIA GR00T 2 humanoids, we outlined the need to distinguish model evaluations from assessments of reproducibility on physical robots and continuous operation. The distinction between simulation training and real-world deployment, covered in a report on humanoid control technology using world models, is also relevant to evaluating this tool. The ability to generate and test programs in simulation is not the same as reproducing the same behavior on physical robots amid variations in equipment and field conditions. For industrial robot deployment, an operational consideration is whether technicians can read, modify and verify generated programs, in addition to whether the movements are correct. Assessing the value of the released tool therefore requires examining which robots and simulators it supports, the safety checks for its outputs, and whether its design allows people to intervene when changes or failures occur in the field. The report that it was released as open source alone does not establish its terms of use or the extent to which it can be externally verified. Moreover, when agents link multiple tasks, it is important to determine where failures at each stage are detected and who gives final approval. The article text provided does not confirm the conditions for evaluation on physical robots, the scale of deployment, licensing or the results of safety testing. Key questions going forward include how to measure the gap between simulation and physical robots, how to assign responsibility for verifying generated code, and how extensively the tool can connect with existing robot control systems. See the original article for details.