A robot’s task score may show whether it completed a job, but not whether success came from the AI model or the software coordinating it. A survey posted on arXiv on Oct. 14 maps agentic robotics into four functional domains and 17 overlapping categories, and argues that benchmark reports should say what each system was allowed to use.
An industrial robotic arm works over an artwork. This illustrative photo is not an image of a system evaluated in the survey.
The paper, “Towards Agentic Robots: Model, Data, Environment, and Harness,” is a survey and organizing framework, not a report of a new robot experiment. The authors use “agentic” for robot systems in which an AI agent reasons from a person’s instruction, interacts with the robot’s environment, corrects execution errors and adjusts its strategy. The survey maps the model and supporting data, tools and tests such systems use.
Four domains, with roles that can overlap
The authors group the research into four areas:
- Model: the AI used for robot-facing interpretation and decisions.
- Data: how information for robot learning is collected and curated.
- Environment: the worlds, tasks and tests in which a robot acts.
- Harness: the software that coordinates models, tools and skills, including memory, monitoring and recovery.
The 17 categories are not four sealed boxes or a required design. A capability can involve more than one area. The map is intended to help researchers describe which functions a system contains and where they sit.
Why a success score needs context
The survey points to a problem with comparing only task completion. One system might rely on a stronger vision-language-action (VLA) executor—the model that turns visual and language inputs into robot actions. Another might use an agent to coordinate several weaker skills, gather extra observations and recover from failures. Similar scores could reflect very different capabilities, the authors say. This is an illustrative comparison in the paper, not a result from a new benchmark.
The authors argue that studies should control for, or clearly report, the executor’s capability, what observations the agent can access, and its interaction or retry budget. They also call for reporting costs such as compute, completion time, extra observations and human assistance, so a higher score is not presented without the resources that produced it.
What the framework does not establish
Because the paper is a survey, it does not test its 17-category map against a shared benchmark or show that using the taxonomy improves robot performance. It offers a vocabulary for describing systems and a case for fairer comparisons; it does not rank architectures or settle which design works best.
Sources
The survey paper on arXiv; the authors’ companion repository.
Image credit: Oleg Yunakov, “KUKA industrial robotic arm performing ‘The Beach’ by Roy Nachum - 02,” via Wikimedia Commons, licensed under CC BY-SA 4.0. The image is illustrative and is not from the paper.