Framework Guides Robots Through Open-Ended Tasks
The research-stage proposal pairs foundation models with a sense-reason-act loop, and its authors say agentic robotics still lacks a shared evaluation standard.
Updated on Nov. 17, 2026 in Robotics

Grid Robotics has published a research paper, posted to arXiv on November 13, proposing a framework that uses foundation models to guide robot autonomy on open-ended tasks. The work is research-stage: it is a proposal, not a shipped product.
Why it matters
The framework aims to move robots beyond preprogrammed routines and task-specific policies toward handling open-ended work. The paper also points to a gap in the field, the lack of a shared definition and reliable evaluation framework for agentic robotics, which is why the authors frame evaluation as part of the proposal.
The paper is a framework proposal, so it reports no measured performance figures. It assesses both the foundation model's capabilities and the performance of the full robot system, but the extraction contains no numbers from those assessments.
The players
Grid Robotics
A robotics company whose research group authored a paper on foundation-model-guided robot autonomy.
OpenAI
An AI research company that distributed its robotics work across research and Applied teams, and whose GPT-6 Astra model is cited as an example that could support the approach.
The details
The framework's central unit is called Embodied System 2, which pairs a foundation model, a large general-purpose AI model, with an embodied harness, the software layer that connects the model to a physical robot's sensors and actuators. The loop runs in repeated steps: the robot takes in its surroundings and a user's intent, reasons about what to do, chooses relevant skills or tools, checks the outcome, and adjusts. Because the steps repeat as conditions change, the system is designed to keep correcting rather than execute a single pre-planned sequence.
Timeline
November 13, 2026: Grid Robotics paper posted to arXiv.
November 17, 2026: Article on the paper published by TokenPost.
The Tech Race
The paper sits within the broader push to apply large foundation models to physical robots, where OpenAI's distributed robotics work is a parallel effort. Its stated contribution is the evaluation gap, which it identifies as a field-level problem that competing efforts will also need to address.
The framework is at the research stage, so no robot using it is available to buyers or users. Practical effects would depend on future prototypes and on whether a shared evaluation standard emerges, and the extraction gives no timeline for either.
The takeaway
The trajectory to watch is whether robots can move from single actions toward broader responsibilities, as the paper's primitives-to-roles path sketches. A concrete marker to track is whether the authors or others publish benchmark results that test the proposed evaluation framework on physical robots.
Further reading
For the wider context on machines that act on their own, see the Robotics section.
Source note: This article includes information reported by TokenPost.


