Until recently, getting a robot to do a new job meant training a new policy for it. A position paper from the team at Grid Robotics argues that GPT-6 Astra changed that: within weeks of its release, general models running inside an agent harness matched or beat task-specific policies on many robot benchmarks without extra training. The authors want the field to agree on what this new kind of system is and how to test it.
They call the approach agentic robotics. Instead of one network mapping camera images straight to motor commands, a foundation model sits at the top as a decision maker, which they call Embodied System 2. It reads the user's intent, plans, calls lower-level skills and tools, checks whether each step worked, and changes course if not. Below it sit learned skills such as vision-language-action models and world-action models (System 1), and below those the body's control loop (System 0). The hard part, the paper says, is turning general reasoning into actions that can be verified under physical limits: little time to react, mechanical constraints, changing surroundings and actions that cannot be undone.
The paper's practical contribution is an evaluation framework that scores both the model's own abilities and how well the whole closed-loop system performs. That matters because current results are hard to compare. The authors point out that published RoboDojo evaluations of Astra use different sources of task knowledge and different action budgets, and that one study showed large gains just from adding execution memory while the planner and skills stayed fixed.
The roadmap runs from single actions to what the authors call standing roles: a robot trusted with an ongoing job, tracking goals over time, recovering from failures and learning from experience without forgetting what it already knew. It also covers which lessons from software agents carry over, such as using tools and checking results, and which do not, since a physical mistake is costly.
This connects to other robot work we have covered, such as Gemini Robotics 2.5 and RoboScript's in-context learning. The caveats are plain: this is a preprint, it is not peer reviewed, it reports no new experiments, and the claim of a paradigm shift rests on a few weeks of results from one model family.