A new class of software is quietly changing who—or what—writes the code that controls robots. Researchers at UC Berkeley, led by veteran roboticist Ken Goldberg, have introduced what they call 'agentic robotics': AI systems that break down a task into component skills, generate control programs for each, test them inside NVIDIA Isaac simulations, diagnose failures, and rewrite the code before it ever touches a physical machine. The approach has already left the lab. Ambi Robotics, a Berkeley-founded warehouse automation company, used the method to analyze real parcel-sorting data and push updated robot behaviors that added more than 15,000 sorted packages per robot annually, with the changes now running across 30 percent of its U.S. fleet.
The shift matters because it attacks one of robotics' oldest bottlenecks: the slow, expensive, highly specialized work of hand-coding and tuning robot behavior for every new task and environment. For decades, deploying a robot on a new job meant engineers writing bespoke control logic, testing it on physical hardware, and iterating through costly trial and error. Agentic robotics proposes instead that AI agents do the decomposition, coding, simulated testing, and revision themselves, compressing a process that once took specialist teams weeks into automated cycles that run largely unattended. If it scales beyond Ambi's sorting robots, the implications stretch across warehouses, factories, and eventually homes, where the economics of robot deployment have long been constrained by the cost of programming and validating new behaviors.
From Demonstration to Deployment
The Berkeley team's framework, called Graph-as-Policy, represents a structural departure from how robots have traditionally been programmed. Rather than a human engineer specifying every motion and contingency, an AI agent takes a high-level task, splits it into a graph of discrete skills, and generates candidate control programs for each piece. Those programs are then run through NVIDIA Isaac simulations thousands of times, where the agent can observe failures, identify the likely cause, and automatically revise the code without waiting for a person to intervene.
What makes the approach significant is not the simulation step itself, which robotics researchers have used for years, but the closed loop between generation, testing, and revision happening autonomously. Researchers reported that a Graph-as-Policy program refined entirely through this simulated trial-and-error process worked successfully when transferred to a real robot, suggesting that the sim-to-real gap—long one of robotics' most stubborn challenges—can be narrowed by letting an AI agent iterate far faster than a human team could.
The approach builds on a broader industry push toward what NVIDIA and others now call physical AI or embodied AI, where models trained largely in simulated environments are expected to transfer their capabilities into physical machines. Goldberg's team's contribution is less about the simulation infrastructure and more about inserting an autonomous agent as the author and debugger of the control code itself.
A Real Warehouse, A Measurable Return
Ambi Robotics offers the clearest evidence yet that agentic robotics can translate into commercial value rather than remaining a research curiosity. The company, spun out of Berkeley's robotics program, applied the method to its parcel-sorting robots by feeding real operational data into the agentic pipeline, which then proposed and tested behavioral changes before they were pushed to physical machines. The resulting updates increased each robot's effective capacity by more than 15,000 packages sorted per year.
By November 2026, those changes had been rolled out across 30 percent of Ambi's United States fleet, a scale that moves the technology well past a pilot project. For a logistics operator, a 15,000-package annual gain per robot is a direct, quantifiable return, the kind of metric that procurement teams and operations executives can weigh against the cost of licensing or deploying the software. It is a sharp contrast to many robotics demonstrations that showcase technical novelty without a clear path to measurable throughput gains.
The Ambi deployment also hints at a new maintenance model for commercial robotics fleets: rather than periodic firmware updates written by engineering teams, robots could receive continuously improving behaviors generated by agents analyzing their own operational data. That model, if it holds up across more companies and use cases, would mean robots effectively get better at their jobs over time without requiring a human to redesign their control logic for each improvement.
The Infrastructure Making It Possible
None of this would be feasible without the simulation and physics infrastructure that NVIDIA has been building out for embodied AI. The company's robotics stack now includes the open GR00T foundation models for natural-language and multistep robot tasks, Cosmos world models that generate synthetic training data, the newly released Newton 1.0 physics engine, and the latest versions of Isaac Sim and Isaac Lab for simulation and validation. Together, these tools form the substrate on which agentic systems like Graph-as-Policy can run their generate-test-revise cycles at the speed and scale required to be useful.
The strategic bet behind this stack is that simulation fidelity and throughput, not real-world trial and error, will become the primary lever for reducing the time and cost of deploying new robot behaviors. Isaac Sim 6.0 and Isaac Lab 3.0 are explicitly built to let an agent run far more test iterations in simulated time than would ever be practical on physical hardware, which is precisely the capability that let Berkeley's researchers validate a Graph-as-Policy program before handing it to a real robot.
This also signals a broader industry alignment: companies ranging from warehouse robotics firms to humanoid manufacturers are converging on the same underlying infrastructure, suggesting that future competitive advantage in robotics may hinge less on proprietary hardware and more on who can most efficiently train and validate behaviors inside these shared simulation ecosystems.
We're moving from robots that need an engineer for every new skill to robots whose skills are written, tested, and improved by AI agents working in simulation before a single physical trial. That changes the economics of deployment entirely.
What Still Stands in the Way
Despite the promise, agentic robotics faces real constraints before it can be trusted with higher-stakes tasks. Industrial safety certification is emerging as a strategic bottleneck rather than an afterthought: organizations including TÜV SÜD, Accenture, and NVIDIA are now actively working on certification frameworks for physical AI systems operating around human workers, a sign that regulators and insurers are not yet comfortable simply trusting agent-generated code without independent validation.
Generalization remains the other open question. Ambi's parcel-sorting environment is relatively structured and repetitive, which makes it a favorable early test case. Whether the same generate-test-revise loop can reliably handle the far messier perception and manipulation challenges found in less structured settings—grasping irregularly shaped objects, locating items hidden under piles, or adapting to constantly changing warehouse layouts—is still being worked out across separate but related research efforts.
Sources
- https://vcresearch.berkeley.edu/news/recent-advances-ai-are-creating-paradigm-shift-robotics
- https://roboticsbusinessnews.com/
- https://www.industryevents.com/news/section/robotics
- https://roboticsbusinessnews.com/category/research-ans-development/breakthrough-technologies
- https://techcrunch.com/category/robotics/
- https://robotoverflow.com/en/news/page/16/
- https://www.automate.org/robotics/news
- https://itbrief.news/tag/robotics
- https://blogs.nvidia.com/blog/national-robotics-week-2026/
- https://www.techbriefs.com/topia/sae/tb/manufacturing/automation?start=0
- https://www.hcamag.com/us/specialization/change-management/humanoid-robots-are-starting-real-jobs-how-soon-will-they-replace-real-staff/594971












Leave a Comment