Paper review · November 17, 2026
A Vision for Agentic Robotics: A New Era of Robot Systems
Envisioning Agentic Robotics: A New Era of Robotic Systems
The authors propose a reference architecture, an evaluation framework and a research roadmap for agentic robotics, built around 'Embodied System 2', which combines a foundation model with an Embodied Harness.
Problem
When frontier models such as GPT-6 Astra are given an agent harness, they match or beat task-specific policies on several robot benchmarks without any additional training. However, there is no shared definition of what agentic robotics is. Evaluation conditions also differ from study to study, which makes results hard to compare. For example, even evaluations of the same Astra model differ in where task knowledge comes from and in the action budget.
Method
The framework splits functions into three roles. S2 handles intent interpretation, reasoning, skill and tool orchestration, and outcome verification. S1 consists of callable skills such as VLAs, WAMs, perception and motion planning. S0 is real-time body control. S2 is made up of a foundation model and an Embodied Harness, and the harness has five parts: a context manager, a tool interface, a planner, memory and a verifier. A high-frequency real-time event monitor can interrupt S2 or trigger an immediate safety response in S0. The authors define three execution paths: hierarchical orchestration, direct control and Code as Policy. On this basis they build a nine-dimension evaluation framework that separates model capability from closed-loop system performance, and present a roadmap that divides the scope of delegable work from L1 (executing individual motions) to L4 (performing an ongoing role).
Results
- The authors offer a conceptual framework rather than experimental results. They lay out the S2/S1/S0 role split, the five core components of the harness and the three execution paths.
- The verification stage is defined by four questions: was the request accepted, did execution finish, does the physical effect persist, and was the user's intent satisfied. The authors argue that this should also catch failures that a skill wrongly reports as successes.
- The evaluation framework separates model capability from overall system performance across nine dimensions. It emphasizes independent outcome verification, transparent reporting and evaluation in environments that keep changing.
- As existing evidence, the authors cite the Harness VLA case. With the planner and VLA held fixed, execution memory alone changed performance substantially, and the authors use this as evidence that fair comparison conditions are needed.
Limitations
This is a perspective and roadmap paper, not a paper that validates a new system experimentally. Within the provided text, there are no results that use the proposed evaluation framework to compare real systems quantitatively. Much of the discussion relies on early results from one recently released frontier model (GPT-6 Astra). The text was cut off partway, so the later sections (analysis of evaluation evidence, roadmap details) could not be checked.
Why it matters
Robot development is shifting from training a policy for each task to having a general-purpose model coordinate skills. This paper offers a shared vocabulary and structure for designing and evaluating such systems. Teams building robot agents can use it as a checklist for reviewing harness components and verification design.
Key terms
- Embodied System 2
- The decision-making core of the robot, combining a foundation model with an Embodied Harness. It interprets intent, calls skills and tools, and verifies outcomes to decide the next action.
- Embodied Harness
- The runtime that assembles the model's input context and turns the model's decisions into actual execution. It keeps execution within constraints and makes it observable and verifiable, and consists of context management, a tool interface, a planner, memory and a verifier.
- VLA(Vision-Language-Action)
- A model that takes images and language instructions as input and directly outputs robot actions. In this framework it sits as one of the S1 skills that S2 calls.
- Code as Policy
- An approach in which the model writes robot control code directly and an interpreter runs that code so it acts as a skill.
Original
Authors: Grid Robotics, :, Qingfeng Lan, Duo Li, Siyao Li, Yiheng Lyu, Shijian Lu, Zhenghao Qi, et al. · Posted 2026-11-13 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.