Grid Robotics Paper Proposes a Framework for Long-Term Robot Autonomy
The proposal pairs a foundation model with an embodied harness and calls for a shared definition and reliable evaluation framework for agentic robotics.

A Grid Robotics research paper posted to arXiv Nov. 13 proposes a framework for using foundation models to guide robots through open-ended tasks and evaluate how the complete systems perform. GPT-6 Astra is presented as one example of a model that could support the approach.
The framework’s central unit, called “Embodied System 2,” pairs a foundation model with an embodied harness to form a robot’s decision-making core. The setup is intended to help robots handle tasks beyond preprogrammed routines and task-specific policies.
In the proposed loop, a 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 its approach. The steps repeat as conditions change, rather than ending with the first action.
The paper identifies a key challenge in turning general-purpose reasoning into actions that work within physical limits. Robots must respond in time, respect mechanical constraints and deal with changes in their environment. The proposal centers on connecting model reasoning with actions that can be carried out and checked in those conditions.
The authors propose assessing both the foundation model’s capabilities and the performance of the full robot system. Looking at both levels would cover the model’s reasoning as well as the robot’s closed-loop process of sensing, acting, checking and adapting.
The paper identifies the lack of a shared definition and a reliable evaluation framework for agentic robotics. It also discusses benchmarks as part of the evaluation challenge. The proposed framework is intended to give researchers a way to assess models and complete systems against a more consistent basis.
The research sketches a longer-term path from individual actions, or primitives, toward robots taking on roles with broader responsibilities. It describes systems that can work safely and reliably amid dynamic, uncertain conditions as a goal for that progression.
The proposal follows OpenAI’s move to distribute robotics work across research and Applied teams, with general models such as GPT-6 Astra identified as part of the company’s robotics development effort.