What Agentic Robotics Means

Agentic robotics is a method in which AI coding agents write, test and repair the code that runs a robot. IEEE Spectrum describes it as turning robot programming from a data problem into a coding problem.

Ken Goldberg, a professor of robotics at UC Berkeley, popularized the term with Spencer Huang, Director of Product for Robotics at NVIDIA. IEEE Spectrum discloses that Goldberg also co-founded an industrial robotics company.

Goldberg and Huang are among the authors of a controlled-study preprint that defines agentic robotics as asking coding agents to implement and iteratively improve robot policies. Those policies improve from simulated and real-world physical interactions. A second preprint, the Variational Automation Benchmark for Agentic Robotics (VABAR), describes the agents as pretrained LLM (large language model) and VLM (vision-language model) coding agents.

Imitation learning starts from demonstrations. It feeds human demonstration data into a machine-learning system, and IEEE Spectrum says the resulting algorithms are essentially black boxes that are difficult or impossible for humans to understand. The VABAR authors describe agentic robotics as combining interpretable robot programming with the open-world adaptability of model-free policies.

How the Write-Simulate-Fix Loop Works

Agentic robotics works through a loop that splits a task, gives each sub-task a coding agent, and has each agent write code, test it in simulation and fix what fails. IEEE Spectrum defines a coding agent as an autonomous AI system that can plan, write, test and modify software with minimal human intervention.

The agents draw on existing robotics tools such as object classifiers, pose estimation, task and motion planning, and trajectory generation. They run their code in simulation and change the starting conditions. Then they identify failures, refine the code and iterate until the program succeeds.

Huang told IEEE Spectrum that AI agents can create "slop". NVIDIA counters it with adversarial agent setups, where one agent designs something and another agent tries to find its flaws. For background on simulation, see the guide to industrial robot simulation software.

IEEE Spectrum's grocery-packing example shows how one job splits across agents. Packing groceries into a bag becomes three sub-tasks, each handled by its own coding agent:

  • Perception agent: identifies the groceries and the bag
  • Manipulation agent: grasps each object with the gripper
  • Transport agent: moves each object once it is grasped

How Agentic Robotics Differs from Imitation Learning

Agentic robotics produces readable robot code from simulated trial and error, while imitation learning produces a learned policy from human demonstration data. Model-based hand coding sits in the third column because agentic robotics shares its output format: code a person can read.

Goldberg told IEEE Spectrum that the imitation learning paradigm "has hit a plateau". He said it is promising for home robots, where speed and reliability matter less. For industrial applications, he said, it is very hard to reach the reliability industry requires, and adding more data does not clearly get there.

IEEE Spectrum says engineers programmed robots for decades with model-based methods, writing detailed code on top of a model of the robot's physics and environment. Those hand-tuned systems could run fast and reliably, but they fell out of favor because they need human time and expertise. The article describes agentic robotics as a move back toward human-readable code written from an understanding of how the robot works, with no data and no humans required.

QuestionAgentic RoboticsImitation LearningModel-Based Hand Coding
What it starts fromA task split into sub-tasks, plus existing robotics toolsHuman demonstration dataA detailed model of the robot's physics and environment
What it producesHuman-readable robot codeA learned policy that IEEE Spectrum calls essentially a black boxDetailed code written by engineers
Who can read and fix failuresA person can understand and fix the code, per GoldbergDifficult or impossible for humans to understand, per IEEE SpectrumNot established
Data it needsNo demonstration data, per IEEE Spectrum and GoldbergHuman demonstrations, and Goldberg says more data does not clearly reach industrial reliabilityHuman time and expertise, per IEEE Spectrum
Reliability evidence so farUp to 99 percent success in UC Berkeley testing reported by IEEE Spectrum, not independently verified20 percent baseline success on some tasks in the same reported comparisonCould run fast and reliably, per IEEE Spectrum
Main weaknessDepends on simulations that are hard to build and on trusting agent outputReliability plateau for industrial use, per GoldbergNeeds human time and expertise

Researcher Claims Awaiting Independent Real-World Results

IEEE Spectrum reports that in UC Berkeley's testing, agentic robotics approaches reached success rates of up to 99 percent, matching human hand-engineered performance. Baseline imitation learning reached only 20 percent on some of the same tasks. The article does not state the test conditions, the figures do not appear in either preprint abstract, and we found no independent verification in the sources we read.

Huang told IEEE Spectrum that at Wistron in Taiwan, retasking one robot for a new product takes at least 8 hours today and would take a minute or two with agentic robotics. That is NVIDIA's claim and has not been independently verified. Huang also said some semiconductor partners need parts placed within 0.3 millimeters, which he said NVIDIA's GR00T humanoid development platform cannot do with repeatability today.

The controlled-study preprint ran nine manipulation tasks, 631 experiment configurations and 28,000 coding-agent rollouts in simulation, varying model capability, skill priors, feedback structure and curriculum. Its abstract says the bottleneck is often the step of interpreting available information into the right policy update. A working self-learning loop depends on how well agents attribute failures, reason about spatial and geometric structure, and selectively restructure code.

Torsten Kroeger, chief science officer at Intrinsic, is not directly involved in the research and told IEEE Spectrum that no physics engine can accurately model physical contact with the world. He also raised the alignment problem: a human engineer can review everything, but trust that agents always do the right thing will become an industry conversation. Four questions remain open in the published material:

  • Independent real-world deployment results
  • Whether the 99 versus 20 percent comparison holds beyond the tasks reported
  • The full methods of both preprints, since we drew on their abstracts only
  • How often agent-generated code is safe to run on a real robot

Agentic Robotics for Industrial Manipulation Cells

Huang told IEEE Spectrum the near-term sweet spot is industrial manufacturing and manipulation tasks that need precision. The VABAR authors call this Variational Automation: tasks a robot performs persistently with known variations in object shape and pose, as distinct from fixed automation. Their benchmark task classes are tool sorting, tool hanging, gear meshing, syringe packing, cable insertion and cable routing.

For a cell that runs one fixed task already covered by classic programming, start with the guide to industrial robot programming instead.

The evidence is researcher-reported, the studies we read are simulation-based, and the conditions of the 99 percent result are not stated, so treat any vendor claim as a pilot to verify on real hardware. The readable code is the practical difference for a buyer, because a person can review it and fix it.

Our read would change with independent real-world deployment results or peer-reviewed full papers that confirm the reported success rates. Until then, ask a vendor these five questions before agreeing to a pilot:

  • Which tasks were validated on real hardware, beyond simulation?
  • Under what conditions were the reported success rates measured?
  • Who reviews the generated code before it runs on a robot?
  • How was the simulation matched to this specific cell?
  • What happens at contact-heavy steps, given Kroeger's point that no physics engine models contact accurately?

Bottom Line

Agentic robotics is a credible new way to program repeatable industrial tasks, because the output is readable code a person can review and fix. The evidence is researcher-reported, the studies we read are simulation-based, and the conditions of the 99 percent result are not stated, so treat any vendor claim as a pilot to verify. Read the IEEE Spectrum report and both preprint abstracts.

Take the five vendor questions above into any agentic robotics pilot discussion before you commit hardware time.

FAQs

Is agentic robotics the same as agentic AI?

Agentic robotics applies AI coding agents to the specific job of programming robots. The controlled-study preprint defines it as coding agents implementing and iteratively improving robot policies from simulated and real-world physical interactions.

Does agentic robotics need training data?

No demonstration data is required, according to Goldberg and IEEE Spectrum. The agents improve code by running it in simulation and fixing failures. Huang said the good-quality code it generates could itself become training data for the next model.

Are coding agents safe to run on a real robot?

That is not established, and no published source covered here reports how often generated code is safe on real hardware. Kroeger raised alignment, meaning trust that agents always do the right thing, as a coming industry question. Because the output is readable code, a person can review it before it runs.

How does agentic robotics relate to simulation software?

Simulation is where the agents test, fail and fix their code. Goldberg said simulations are non-trivial to build, and his team had an AI agent tune simulation parameters to match a real-world video. IEEE Spectrum says benchmarks are being developed with NVIDIA Isaac Lab-Arena and the Newton physics engine.

Does agentic robotics replace imitation learning?

Goldberg did not claim that to IEEE Spectrum: "I don't want to say that we have the solution". Huang's thesis is that general intelligence for robotics will first come as an agent orchestrating conventional technologies and specialized policies.

Primary Sources