UC Berkeley and NVIDIA propose agentic robotics as an alternative to data-hungry training
At the 2026 Conference on Robot Learning, UC Berkeley and NVIDIA presented a way to program robots through AI coding agents instead of huge sets of demonstrations.
UC Berkeley and NVIDIA announced a new method called agentic robotics at the 2026 Conference on Robot Learning. Instead of training a robot on large piles of human demonstrations, the method asks AI coding agents to write readable control code for industrial tasks. A person describes the job in plain language, and the job is split into smaller parts. In the Berkeley example, a robotic arm packs groceries into a bag, with separate agents handling perception, grasping and moving objects.
The agents test their code in simulation, adjust starting conditions, find failures and repeat until the program works. Berkeley professor Ken Goldberg says an agent was also given access to simulation settings and asked to match a real video. It reproduced the lighting and background closely enough that the two were hard to tell apart.
Why it matters
Goldberg argues that learning from demonstrations has plateaued for industrial work, where reliability matters most. Code that humans can read is also easier to check than a black-box model. Still, Torsten Kroeger of Alphabet's Intrinsic, who was not part of the work, warns that no physics engine models real contact accurately, so the gap between simulation and reality remains open. The report gives no success rates or real-world test results, so the next thing to watch is how well this code performs on physical robots in factories.
