Towards Agentic Robots: Model, Data, Environment, and Harness

Towards Agentic Robots. Model, Data, Environment, and Harness. By Chonghan Qin and fifteen others, from the Chinese University of Hong Kong, Fudan University, and four other institutions. Published on the fourteenth of October. Its code repository is public but slight: three source files.

Foundation models are shifting what robotic intelligence means. The focus is moving beyond predicting low-level actions, toward goal-directed orchestration of the whole pipeline: collecting data, training models, planning tasks, and executing in the real world. The authors call this emerging paradigm agentic robots. In it, agents reason over human instructions, interact with robotic environments, correct their own execution errors autonomously, and adjust their operational strategies as they go.

The field is moving fast, but its studies are not well connected. There is no consistent view among researchers of how to combine the different agentic functions into a complete robot system. That missing picture is what this survey sets out to supply, as a unified overview of existing research on agentic robots.

The overview proceeds in three steps. It first traces the path from robot policies to agentic systems, covering vision-language-action models, world models, and world-action models. It then pins down a definition and scope for agentic robots. And it organizes existing research into seventeen categories, grouped under four functional domains: Model, Data, Environment, and Harness.

Each domain answers a different question about what an agent does for a robot. Under Model, the question is how agents shape model learning and robot decisions. Under Data, how agents collect and curate robot data. Under Environment, how they construct worlds, tasks, and tests, for learning and for evaluation. And under Harness, how capabilities are organized for robot use.

The thread running through all four is a change in the agent's role. Rather than being the policy that emits the next action, the agent becomes the orchestrator around the robot: the one that gathers what the robot learns from, builds where it is trained and tested, decides how it is shaped, and assembles its capabilities into something usable. Read this way, the four domains are four places where that orchestration happens, and the seventeen categories are the finer divisions of each.

The survey ends by discussing open challenges and outlining directions for future research. Its stated hope is a clearer understanding of agentic robots, and support for new approaches that make them more capable, more reliable, and more adaptive. For a listener entering the area, the four-domain map is the thing to carry away: model, data, environment, and harness, as the places to look when asking what an agent contributes to a robot.