Agentic robotics adds 15,000+ packages per robot each year

Why has data held robotics back?

Two traditional methods make very different demands:

• Model-based robotics: Engineers directly program how a robot perceives objects, plans movements and controls actions. This route demands extensive engineering.
• Model-free robotics: Robots learn behaviors from demonstrations through imitation learning, meaning they copy human examples. This route demands enormous amounts of data and time.

There is no easy way to collect enough physical information for robots. In 2025, Goldberg warned of a 100,000-year data gap, comparing robot learning with AI agents that gained language fluency over several years from vast internet text and images, for which robotics has no parallel.

Goldberg said the collection burden had discouraged him. He called agentic robotics a missing link between the two methods and said it offered a way forward that, in his view, did not require additional data.

How does Graph-as-Policy work?

Earlier in 2026, Goldberg’s team reorganized robot code into a framework called Graph-as-Policy. It divides a larger task into individual skills:

• Finding a target object in the scene
• Planning how to grasp the target
• Moving the gripper without hitting anything

Goldberg likened the skills to building blocks that connect into a larger robot program. AI agents assemble them, use NVIDIA’s Isaac libraries to rehearse the program under task variations, identify failures and refine it before deployment.

OpenAI’s recently released Astra model made the workflow more promising, Goldberg said. Postdoctoral students Eric Kaiyuan Chen and Shuangyu Xie gave Astra access to their simulation interface and a short video of a real robot using a sponge to wipe ink from a metal bar.

In less than an hour, the AI tool created and fine-tuned a physics simulation that accurately reproduced the video, including realistic changes in the sponge’s shape. The researchers then combined that simulation with Graph-as-Policy, allowing the AI to generate a control program and repeatedly test and improve it.

When the researchers ran the resulting code on the real robot, it worked. In one trial, however, the robot pushed over the metal bar because it did not reliably infer the bar’s mass.

The failed attempt supplied new evidence: video let the agent improve the simulation, revise the program and test another approach. Goldberg said this was when he got goosebumps. The episode shows why virtual rehearsals still require checks against physical reality.

What has changed in industrial warehouses?

Earlier in 2026, Berkeley-based Ambi Robotics, which specializes in logistics and warehouse order fulfillment, applied the methods to parcel-sorting data and tested updates to robot behavior. The company deployed the change across 30% of its U.S. fleet after tests showed an increase of more than 15,000 packages sorted annually per robot.

“AI-powered industrial robots must continuously improve reliability and performance,” said Jeff Mahler, Ambi Robotics’ co-founder and chief technology officer. “Agentic robotics converted production data into robot behavior updates, increasing throughput across our U.S. fleet while keeping engineers in control.”

Goldberg said the implications are most promising for repetitive industrial tasks performed at scale. He called the ability to self-improve key parts of a task a major step, while cautioning against expecting home assistants to fold laundry yet.

Goldberg said agentic robotics combines classical control theory—the engineering discipline that uses feedback to control machines—with physics and rapidly evolving agentic coding. He called the combination fast and practical for real engineering applications, and said his sense was that it represents a true paradigm shift.

In a 40-year robotics career, Goldberg found only two moments that stirred similar excitement. He named the public release of the World Wide Web in 1993 and ChatGPT’s debut in 2022.