Agentic robotics
Readable robot programs
A UC Berkeley lab has AI agents write readable robot programs, which scored below human-written ones in all twelve of its own test conditions on one simulator.
Robots · Checked 13 Nov 2026 · 13 sources · 10 min

A robot arm on a factory line does nothing until someone has written out its job. Take a job as small as wiping ink off a metal bar with a sponge. For the arm, an engineer has to work out and write down where the bar is, where the sponge goes down, how hard to press, which way to move and when to stop. Then the engineer runs the program, watches what goes wrong, changes it and runs it again until the arm gets it right. What is left is a program a trained person can open and read.
The other established way skips the writing. A large AI model is trained on large datasets of physical demonstrations, recordings of a job being carried out for real, until it can produce the movements itself. The recordings are needed in bulk, and what the model learns is held as a mass of numbers, with no steps in it for anyone to read.
On 9 November 2026 UC Berkeley's newsroom reported that Ken Goldberg is presenting "agentic robotics" at the Conference on Robot Learning. The work comes from UC Berkeley's AUTOLab, with co-authors at NVIDIA, Bosch and Siemens. It hands the engineer's loop of writing, running and changing to software: teams of AI agents write the robot's program, test it in simulation and revise it. The starting material is one human demonstration of the job, and what comes out, in IEEE Spectrum's phrase, is "human-readable code". By the team's description, none of the most advanced AI models runs in the fast loop that steers the arm.
An AI agent is a language model, the kind of software behind a chatbot, connected to tools so that it can act and then see what its action did. Given a programming job, it writes a file, runs it, reads the error and tries again, much as the engineer did with the arm. But a robot program runs on a robot, and an arm running a first draft can break things, so the agents need somewhere safe to try one.
That place is a simulation: a program that holds a copy of the scene, and rules of physics for working out what happens when something in the copy moves. The site of a project called RoboSimStudio says a Claude or Codex agent (both commercial AI products) builds a simulation's scene, camera and physics from a smartphone video and a one-line prompt in about an hour, with humans and agents tuning the copy to match the video. Berkeley News credits OpenAI's Astra with one such build, of a robot wiping ink off a metal bar with a sponge.
With a simulated bar to practise on, the agents can write, and they don't start from a blank page. The team's system, Graph-as-Policy or GaP, is described in its paper as a multi-agent coding harness: a framework in which several agents build a program from a library of ready-made parts called skills. The library starts with 51 skills. A program built from it is a set of nodes, and the paper names three kinds: perception, planning and control. Arrows between nodes show whose output goes to whom, and the whole diagram is called a graph.
Suppose the sponge job were written this way, in four nodes: a perception node that finds the bar in the camera image, a planning node that lays out a path along it, and two control nodes, one to press the sponge down and one to wipe along the path. These four are invented here to show the form, and are not the team's. The project page says every arrow is checked before the program is ever run, to see that what one node sends is the kind of thing the next accepts. In the invented graph, that check would stop the pressing node from being handed a camera image where it expects a path.
Then the graph is rehearsed. It runs in the simulation again and again, failures are traced to specific nodes, and the agents edit the graph and rehearse it once more, until the score stops rising. Suppose the simulated arm tries the invented graph 10 times, with the bar set down a little differently each time, and wipes it clean in 7, a score of 70%. Suppose all three failures are traced to the node that presses down. The agents would rewrite that node and rehearse again, and might then get 9 clean runs in 10, or 90%. When a further round brought no gain, the loop would end. These counts are made up along with the graph, and none of them is a result the team reported.
One trial pushed the bar over
Nine clean runs in ten, in the supposed case, is a score against a simulated bar. The last stage moves the program to a real arm. Of the team's actual sponge program, Berkeley News says that it worked on a real robot, though in one trial the robot pushed the bar over. No source gives the number of trials.
A program revised in rehearsal has been fitted to the rehearsal. In the supposed case the agents rewrote the pressing node until the simulated bar came clean. The judge of each rewrite was the simulation's version of how a sponge and a bar behave when pressed together, and wherever the real pair behaves differently, the rewrite was tuned to the wrong thing. Torsten Kroeger of Alphabet's Intrinsic, the only outside expert quoted in independent coverage, told IEEE Spectrum that no physics engine can accurately model physical contact, a physics engine being the part of a simulation that works out how things move and what happens when they touch. Wiping with a sponge is contact from start to finish. The team's overview says physics models remain imperfect, and a case study called Flip Anything, with Goldberg among its authors, reports that real runs fail in ways the simulator rarely produced.
The team's own benchmark shows, in one case, how far a score can depend on the simulator. A benchmark is a fixed set of tasks with a fixed way of scoring, so that different attempts can be compared. This one, called VABAR, has six task classes at two difficulty levels, twelve combinations in all, with 30 instances per level. Its site describes the kind of work it covers as a robot performing varying instances of one task, with known variation in object shape and pose, pose being where an object sits and which way it is turned. VABAR ships on two simulators, Newton and Isaac Lab-Arena. In its paper, on a task called Syringe Packing at level 1, one graph written by a person scored 87.0% on Newton and 52.5% on Isaac Lab-Arena. The graph was the same in both runs and the simulator was not, so part of any score earned in rehearsal belongs to the simulator.
VABAR, Syringe Packing at level 1, one graph written by a person
The same graph on two simulators
GaP's counts from real robots are also the authors' own. On three tasks its project page reports successes that come to 71 of 75 trials, against 18 of 75 for a comparison system called TipTop. The GaP paper says reliability is not yet at industry levels and that its speed is well below an industry standard of 500 units per hour. The team's overview adds that coding agents are slow, and that whether the approach scales to generalist robots or humanoids is not yet known.
Behind human-written graphs in all twelve
The VABAR paper sets graphs written by agents against graphs written by people. On Newton, graphs that GaP generated with a model called Claude Opus 5 stayed below human-written graphs in all twelve conditions. On Syringe Packing at level 2, for example, the agents' graph scored 12.1% and the human-written one 63.0%. The same paper sets GaP against Claude Code, a coding agent running the same model without the harness or the skill library, and GaP's graphs came out ahead in 11 of 12 conditions.
VABAR on Newton, Syringe Packing at level 2
A graph written by agents against one written by a person
The benchmark's website reports a separate comparison. Eight AI models that write code, each given the same skill library and two rounds of feedback, reached final scores from 44 to 77 on five tasks, against 82 for the human-authored program, which several of them beat on USB-C insertion and cable routing. The procedures differ, so the two studies can't be compared directly.
Every one of these performance figures is the authors' own, and that limits what they settle. Whoever runs a test decides which tasks are tried, how many times, and what counts as success. Those choices can be made in good faith and still shape the answer, so a result gains weight when a stranger makes them afresh and gets the same one. No repeat of these tests by an outside group, and no outside run of VABAR, could be found, though the means to attempt one are public. The code is public under the Apache-2.0 license, which lets anyone use and change it, and VABAR's instances are drawn under a controlled seed, a setting that makes a re-run reproduce the same scenes.
One result comes from a working business. Ambi Robotics, a company Goldberg co-founded, has robots that sort packages. Its blog says agents studied 500 production events and proposed three changes within 10 hours. The changes then went through 10 days of A/B tests, the name for trials that run a changed version beside the unchanged one. Two of the three lowered the estimated number of sorts per robot per year, by 374 and by 1,497. The third raised it by 15,725, a figure extrapolated from 4.2 more packages per hour, and Berkeley News reports that the change was deployed across 30% of Ambi's U.S. fleet. Ambi measured the result itself and does not say how many robots were tested.
An essay by Goldberg puts what would change as a conditional: if these capabilities continue to improve, the bottleneck in robotics may shift from collecting robot data or hand-tuning each application to specifying objectives, constraints, interfaces and verification criteria for the agents. That is a move from writing the program to stating exactly what the job is and how to tell it was done right. A finished GaP program is a JSON workflow file plus scripts, JSON being a plain-text format, so an engineer can read and edit what an agent wrote. Whether engineers will review it at production scale is the trust question Kroeger raised in IEEE Spectrum, when he asked how anyone can know the agents are doing the right thing.
Check it yourself
- Open the Berkeley News feature and find the passage on the sponge and the metal bar. The feature confirms the date and the "is presenting" wording, and that passage confirms the trial in which the robot pushed the bar over.
- Open the IEEE Spectrum report and find Torsten Kroeger's remarks. The report confirms his statement that no physics engine can accurately model physical contact, the phrase "human-readable code" and that Goldberg co-founded Ambi Robotics.
- Read the GaP project page and its result tables. They confirm the library that starts with 51 skills and the real-robot counts, which come to 71 of 75 trials against 18 of 75.
- Read the hypothesis table in Ambi's blog. It confirms that two of the three changes the agents proposed lowered the estimated number of sorts.
- Find the eight-model comparison on the VABAR project website. It confirms the final scores from 44 to 77, against 82 for the human-authored program.
- Compare the human-written Newton column with the GaP column in Table II of the VABAR paper. It confirms the gap in all twelve conditions.
- Follow the instructions in the VABAR repository on a Linux machine with an NVIDIA graphics processor and run one cell of Table II with the paper configuration. Comparing the result with the table tests the authors' figures directly.
Sources
- Agentic Robotics: A new approach to controlling robots. UC Berkeley AUTOLab.
- Recent advances in AI are creating a 'paradigm shift' for robotics. Berkeley News, 9 November 2026.
- Report on agentic robotics, by Evan Ackerman. IEEE Spectrum, 9 November 2026.
- RoboSimStudio project site. RoboSimStudio authors.
- GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks. arXiv, 6 July 2026, revised 4 November 2026.
- GaP project page. GaP authors.
- Flip Anything, project page. Xie, Chen and Goldberg.
- VABAR repository. GitHub.
- VABAR project website. VABAR authors.
- VABAR: A Variational Automation Benchmark for Agentic Robotics, full text, version 1. arXiv, 4 November 2026.
- GaP: A Graph-as-Policy Multi-Agent Self-Learning Harness For Variational Automation Tasks, full text, version 3. arXiv, 4 November 2026.
- Ambi Robotics blog post on agentic robotics. Ambi Robotics, 17 September 2026.
- Goosebumps: a Paradigm Shift is Occurring in Robotics. Ken Goldberg on X, 18 September 2026.