Introduction
The robotics industry has placed a massive bet on data. The prevailing idea is that if artificial intelligence models consume enough examples of human behavior, machines will eventually handle nearly any manual task. Yet a growing number of researchers warn that this data-first strategy may be reaching its limits. At the 2026 Conference on Robot Learning, teams from UC Berkeley and Nvidia are proposing a different path: agentic robotics, which treats robot programming as a coding challenge rather than a data-collection problem.
What Happened
UC Berkeley and Nvidia jointly announced agentic robotics at the 2026 Conference on Robot Learning. The approach uses large language models and autonomous coding agents to generate human-readable software for robots. Instead of gathering thousands of hours of human demonstrations, engineers can describe a task in natural language. Coding agents then break the job into subtasks—such as perceiving objects, gripping them, and moving them—and write, test, and refine code in simulation until it works.
Ken Goldberg, a robotics professor at UC Berkeley and cofounder of Ambi Robotics, explained that imitation learning has reached a performance ceiling in industrial settings. While data-driven methods may work well enough for home robots where occasional errors are tolerable, factories require far greater reliability. Goldberg and Spencer Huang, Nvidia's director of product for robotics, say recent advances in agentic coding have made it viable to let AI write robot code directly.
In testing, the Berkeley team reported success rates as high as 99 percent on certain tasks, compared with roughly 20 percent for baseline imitation-learning methods. The researchers also demonstrated that an AI agent could align a physics simulation with real-world video, matching lighting, background, and environmental details automatically.
Why This Matters
Data-hungry AI models often function as black boxes. Even when they perform well, it can be difficult or impossible to understand why a robot made a specific decision. That opacity is a serious concern in commercial and industrial settings where safety and predictability are non-negotiable. Agentic robotics offers code that humans can inspect, audit, and fix.
The method also addresses a practical bottleneck: there is no internet-scale dataset for robotics comparable to what exists for language models. Collecting robot training data is slow and expensive. By shifting the problem to code generation, teams can tap into the broad knowledge already embedded in large language models.
Early applications are already emerging. Some semiconductor partners require component placement accurate to within 0.3 millimeters, a level of precision that current humanoid platforms cannot reliably achieve. At a Wistron facility in Taiwan, switching a robot to a new product once required a minimum of eight hours of retraining; the new approach reduces that to roughly a minute or two.
Not everyone is fully convinced. Torsten Kroeger, chief science officer at Alphabet's Intrinsic, cautioned that no physics engine perfectly models physical contact, and he raised the alignment problem: how do we know autonomous agents will always make safe and secure decisions? Human review remains essential, especially as trust in these systems grows.
Key Takeaways
- UC Berkeley and Nvidia introduced agentic robotics at the 2026 Conference on Robot Learning.
- The approach uses AI coding agents to write robot software from natural language instructions, reducing dependence on large demonstration datasets.
- Berkeley tests showed success rates up to 99 percent, versus about 20 percent for standard imitation-learning baselines.
- Generated code is human-readable, making it easier to audit, fix, and certify for industrial safety.
- Early use cases include semiconductor manufacturing and rapid robot retasking in electronics assembly.
- Experts still warn about simulation-to-reality gaps and the need for human oversight of autonomous agents.
Conclusion
Agentic robotics may not replace data-driven AI entirely, but it offers a compelling alternative for tasks where precision, explainability, and speed matter most. By turning robot programming into a coding problem, researchers hope to bypass the data-scaling wall that has slowed progress toward general-purpose machines. The technique shows promise, but real-world deployment with industry partners is the next critical step. Whether this becomes the foundation for robot general intelligence remains an open question, yet the shift from data collection to code generation is already changing how some companies think about automation.

Discussion
Join the conversation
Thoughtful reactions, questions, and follow-up ideas help shape the next story.