AI Agents Are Now Writing Robot Code: Berkeley and NVIDIA’s ‘Agentic Robotics’ Already Lifted a Warehouse Fleet’s Output

For most of the last decade, getting a robot to do something new came down to two expensive options. Engineers could program every perception, planning and control step by hand, or they could feed the robot huge piles of demonstration data and hope it learned the behavior. UC Berkeley roboticist Ken Goldberg now says there is a third path, and he calls it a “paradigm shift.” In a Berkeley News report published November 9, 2026, Goldberg and NVIDIA’s Spencer Huang describe “agentic robotics,” an approach where teams of AI agents write, test and repair robot programs in simulation before those programs ever touch real hardware.

What Agentic Robotics Actually Means

The idea borrows directly from what AI agents have already done to software development. Instead of a single model trying to learn an entire task end to end, the work is split into reusable skills, such as finding an object, planning a grasp and moving a gripper without collisions. Goldberg’s team calls this structure Graph-as-Policy and compares the skills to building blocks.

AI agents then assemble those blocks into a robot program, run it through many task variations using NVIDIA’s Isaac simulation libraries, find where it fails and revise the code. Because the output is readable code rather than an opaque neural network, human engineers can inspect what the robot is doing, fine-tune it and verify it. Goldberg told Berkeley News that the approach “provides a missing link between model-based and model-free methods” and is “a way forward that doesn’t require any additional data.”

That last point matters. Goldberg has previously warned of a “100,000-year data gap” holding robotics back, because there is no robot equivalent of the vast internet text that trained today’s language models. Agentic robotics sidesteps that gap by leaning on simulation and code generation instead of ever-larger demonstration datasets.

From a Sponge Test to a Real Warehouse Fleet

The Berkeley team’s most striking example involved a robot wiping ink off a metal bar with a sponge. According to Berkeley News, postdoctoral researchers Eric Kaiyuan Chen and Shuangyu Xie gave OpenAI’s Astra model access to their simulation interface and a short video of the real robot. In under an hour, the AI built and tuned a physics simulation that reproduced the video, including realistic sponge deformation. It then generated a control program, improved it in simulation, and the resulting code worked on the real robot.

It was not flawless, and that is part of the story. In one trial the robot knocked the bar over because it misjudged the bar’s mass. The agent used video of the failure to improve its simulation, revise the program and try again. “This is when I got goosebumps,” Goldberg said.

The business proof point comes from Ambi Robotics, a Berkeley-based company focused on warehouse order fulfillment. Ambi applied the method earlier this year to analyze parcel-sorting data and test updates to robot behavior. Berkeley News reports the tests showed capacity rising by more than 15,000 packages sorted per robot per year, and the change is now deployed across 30% of Ambi’s U.S. fleet. Ambi co-founder and CTO Jeff Mahler said agentic robotics “converted production data into robot behavior updates, increasing throughput across our U.S. fleet while keeping engineers in control.”

Open Code and a New Industrial Benchmark

Goldberg is presenting the results this week at the Conference on Robot Learning (CoRL) in Austin, Texas. His team is releasing open-source code built on NVIDIA’s Isaac framework, plus a benchmark for measuring robot performance across industrial tasks. NVIDIA’s own CoRL announcement frames agentic robotics as “a new paradigm that integrates frontier agents to create interpretable code for physical robotics systems,” and notes that the conference’s focus is shifting from locomotion toward delicate manipulation such as aligning gears and connecting cables.

Huang summed up the problem the partnership is targeting: robots are advancing quickly, he said, but developing them for real-world environments “remains slow, data-intensive and difficult to scale.”

Why It Matters for Business

Most companies that buy automation do not struggle with the robot hardware itself. They struggle with integration: the months of engineering time it takes to get a robot reliably doing one specific job, and the cost of re-engineering it every time a product, package size or line layout changes. Agentic robotics goes straight at that bottleneck.

• Faster deployment and updates. If agents can draft and rehearse robot programs in simulation, the time from “new task” to “running on the floor” could shrink from months to days for some jobs.
• Robots that improve from their own production data. Ambi’s results suggest a fleet can get measurably more productive after it is installed, which changes the return-on-investment math for automation buyers.
• Auditability. Because the output is interpretable code, operations and safety teams can review what changed, a big advantage over black-box models in regulated or safety-sensitive settings.
• Lower data costs. Companies that cannot afford to collect massive demonstration datasets may still be able to automate, using simulation and agents instead.

Leaders should keep expectations grounded. Goldberg himself says the near-term payoff is in industrial settings where tasks are repetitive and run at scale, and he cautioned not to expect home robots folding laundry yet. Simulation never perfectly matches reality, as the toppled metal bar showed, so human oversight and real-world testing remain essential. The most practical early adopters are likely to be logistics, warehousing and light manufacturing operators that already run robot fleets and generate the production data these agents can learn from.

The Bottom Line

Goldberg says only two other moments in his 40-year career felt like this one: the public release of the World Wide Web in 1993 and ChatGPT’s debut in 2022. Whether agentic robotics lives up to that comparison will depend on how well it scales beyond research labs and early adopters like Ambi. But the direction is clear. The same AI agents that rewrote how software gets built are starting to rewrite how robots get programmed, and businesses planning automation investments should factor that into their roadmaps now.

Sources: Berkeley News, NVIDIA Blog.

Edited by Coron @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous