skulldc

A position paper names Embodied System 2 as the decision core of agentic robotics

A 13-author paper argues that a foundation model paired with an embodied harness should drive robot decisions, and proposes an evaluation framework plus a research roadmap toward sustained autonomy.

For years, the usual way to get a robot to handle a family of tasks was to train a dedicated policy for it. A paper posted to arXiv on 2026-11-13, with 13 authors and Grid Robotics listed first, argues that this is giving way to what it calls agentic robotics: applying a foundation model to open-world embodied tasks instead of training one policy per task family.

The claim it starts from

The abstract opens with the release of GPT-6 Astra, which it presents as the point where agentic robotics emerged as a new paradigm. It states that within weeks, frontier models with agentic harnesses matched or exceeded task-specific policies on many embodied benchmarks. The abstract does not list those benchmarks or the systems compared, so that sentence reads as framing by the authors rather than as a result measured in the paper.

An agentic robotic system, in their description, runs in a closed loop: it observes the environment, interprets user intent, reasons about the task, invokes skills and tools, verifies outcomes and adapts its strategy. The authors note that the community has not yet converged on a shared definition, a reliable evaluation framework or a research roadmap, and the paper sets out to offer one of each.

Embodied System 2

The central argument is that Embodied System 2, a foundation model together with an embodied harness, forms the decision-making core of agentic robotics. The central challenge, as the abstract puts it, is translating general-purpose reasoning into actions that are effective and verifiable under physical constraints: limited reaction time, mechanical limits, unpredictable changes in the environment and irreversible consequences. That last item is why verifying outcomes sits inside the loop the authors describe, not after it.

Evaluation and roadmap

From that perspective the authors propose an evaluation framework that measures two things: the capabilities of the model and the performance of the closed-loop system. They also identify principles that carry over from digital agents to physical ones, and the challenges that must be addressed before physical autonomy becomes reliable.

The roadmap is built on top of Embodied System 2 and the evaluation framework. Its target is sustained autonomy, which the paper defines as the ability to perform work safely and reliably with progressively broader responsibility, moving from executing individual primitives to fulfilling standing roles.

What to keep in mind

This is a perspective and agenda paper, not a report of a new model or a new benchmark score. The abstract gives no metrics for the proposed framework and does not say whether it ships with code or a public test suite, and the Hugging Face page lists no code repository for it. Readers looking for the concrete evaluation protocol, and for how closed-loop performance is scored, will need the full text.


Source: Envisioning Agentic Robotics: A New Era of Robotic Systems (arXiv 2611.18359)