Researchers frame agentic robotics around foundation models acting in closed loop
An arXiv position paper argues that a foundation model plus an embodied harness can serve as the decision-making core of physically autonomous robots, and proposes an evaluation framework.
In plain words
A research paper describes "agentic robotics": instead of training a robot a separate policy for each job, one general-purpose AI model runs in a loop, watching the environment, interpreting what a user wants, calling skills and tools, and checking results. The authors argue this loop needs a way to turn general reasoning into safe physical actions, and they propose a framework for measuring it.
What was announced
The paper, an arXiv abstract, opens by citing the release of GPT-6 Astra as marking what the authors call a new paradigm, noting that within weeks frontier models paired with agentic harnesses matched or exceeded task-specific policies on many embodied benchmarks. Rather than training a dedicated policy per task family, an agentic robotic system applies a foundation model to open-world embodied tasks, operating in a closed loop: observing the environment, interpreting user intent, reasoning, invoking skills and tools, verifying outcomes, and adapting strategy.
The authors argue that what they term Embodied System 2, a foundation model combined with an embodied harness, forms the decision-making core of the field. Its central challenge is translating general-purpose reasoning into effective, verifiable actions under physical constraints such as limited reaction time, mechanical limits, unpredictable environment changes, and irreversible consequences. They propose an evaluation framework measuring both model capabilities and closed-loop system performance, and sketch a roadmap toward sustained autonomy, from executing individual primitives to fulfilling standing roles. The abstract notes the community has not yet converged on a shared definition, reliable evaluation framework, or research roadmap.
What it does not show
- Only the paper's abstract was reviewed, so the full article's methods, experiments, and details are unverified.
- The claims are the authors' position and reported observations, not independent confirmation of benchmark results or of GPT-6 Astra's capabilities.
- The abstract does not report empirical results from the proposed evaluation framework, so its effectiveness is not established.
Sources
Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 17 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.