Glossary
November 9, 2026

AI Programs Write the Instructions for Factory Robots and Rehearse Them in a Simulation First, and a Parcel-Sorting Company Is Already Using It

Researchers at UC Berkeley and the chipmaker Nvidia have shown a way for a team of AI programs to write the instructions a robot follows, practice them inside a computer simulation, and fix what goes wrong before the robot ever moves. A company that makes parcel-sorting robots says it is already using the method on almost a third of its U.S. robots. The figures come from the researchers and the company, with no outside check yet.

Researchers at the University of California, Berkeley, and Nvidia, the company best known for the computer chips that power AI, today described a new way to get robots to do jobs in factories and warehouses. Instead of a person writing every instruction by hand, a team of AI programs writes them, tests them, and rewrites them until they work.

Teaching a robot a new job is usually slow. One common way is to have people show the robot what to do thousands of times so it can learn by copying, which takes huge amounts of recorded examples. The other is to have skilled engineers write the robot's instructions step by step, which takes a long time and breaks easily when a part arrives at a slightly different angle.

In the new approach, the job is broken into small, ready-made skills, such as finding an object, planning how to grip it, and moving the arm without bumping into anything. AI programs then string those skills together into a plan for the robot. Before anything happens in the real world, they try the plan over and over inside a simulation, a detailed computer model of the robot and the objects around it, much like a flight simulator for pilots. When the plan fails in the simulation, the AI programs change it and try again. Human engineers check the result before it goes onto a real machine.

In one example, the researchers gave an OpenAI AI program called Astra a short video of a robot wiping ink off a metal bar with a sponge. Within about an hour, it built a simulation that matched the video, then wrote instructions that worked on a real robot. It did not go perfectly: in one try the robot knocked the bar over because the program had guessed its weight wrong. The researchers fed the video of that mistake back in to improve both the simulation and the instructions.

The method is already in use outside the lab. Ambi Robotics, a Berkeley company whose robots sort parcels for shipping, used it to study its sorting records and test changes to how its robots behave. Berkeley says the change lets each robot sort more than 15,000 extra packages a year, and that it now runs on 30% of the company's robots in the United States. "Agentic robotics is a fundamental shift in how robots are controlled," said Ken Goldberg, the Berkeley professor who leads the work, using a term for AI programs that take a series of steps on their own.

There are limits. The parcel figures come from Berkeley and Ambi, with no details of how they were measured and no outside check. The research paper is being presented at a robotics conference this week, and no independent experts have weighed in yet. Goldberg says the approach suits repetitive factory and warehouse work, and that household chores like folding laundry are still out of reach. The team is releasing its code for free so others can test it.

Sources

  1. UC Berkeley — news release on agentic robotics (via EurekAlert)eurekalert.org · primary
  2. arXiv — GaP: A Graph-as-Policy multi-agent self-learning harness (research paper)arxiv.org
  3. The Robot Report — can AI agents program industrial robots without training data?therobotreport.com
Follow the timeline
Get an email when new entries are added.
© 2026 Sugarpine