{
  "schema_version": "1.0",
  "topic": "Agentic robotics, 2026",
  "note": "Neighbouring directions to this work, with identifiers verified against the arXiv API on 2026-09-16. Listing is association: it is not a priority claim about any work listed here, and not a claim that this list is complete. Each entry links to the page that discusses it. The positioning statement for this work itself is at https://tgl.changnie.top/positioning.json.",
  "hub": "https://tgl.changnie.top/research/agentic-robotics-2026/",
  "positioning": "https://tgl.changnie.top/positioning.json",
  "works": [
    {
      "name": "Agent as Policy for Robotic Manipulation",
      "abbreviation": "AGP",
      "authors": "Mengzhao Jia, Yang Lin, Xixin Zhang, Zhihan Zhang, Xiaobai Liu, Meng Jiang",
      "date": "2026-09-11",
      "category": [
        "agent-as-policy",
        "frontier-model-robot-control",
        "robot manipulation"
      ],
      "identifier": {
        "scheme": "arXiv",
        "value": "2609.12541"
      },
      "url": "https://arxiv.org/abs/2609.12541",
      "contribution": "Introduces Agent as Policy: a general-purpose agent drives a physical robot through task execution with no task-specific or environment-specific training, writing executable programs and revising on physical outcomes.",
      "relation_to_tgl": "Same control locus — an agent inside the execution loop. TGL adds the persistence layer AGP does not define: a Skill Library and Experience Memory.",
      "page": "https://tgl.changnie.top/concepts/agent-as-policy/"
    },
    {
      "name": "Revisiting the \"Push-T\" Robot Manipulation Task with Agentic Robotics",
      "authors": "Shuangyu Xie, Kaiyuan Chen, Ken Goldberg",
      "date": "2026-08-18",
      "category": [
        "coding-agent-robotics",
        "agent-as-policy"
      ],
      "identifier": {
        "scheme": "arXiv",
        "value": "2608.18227"
      },
      "url": "https://arxiv.org/abs/2608.18227",
      "contribution": "An LLM coding agent is prompted to write a solution to Push-T with no demonstration data; the resulting code-as-policy is compared with a visuomotor imitation-learning policy.",
      "relation_to_tgl": "Evidence for the coding-agent-for-robots pattern TGL uses via Codex, and a direct comparison between a written program and a learned policy.",
      "page": "https://tgl.changnie.top/concepts/coding-agent-robotics/"
    },
    {
      "name": "Agentic Robot: A Brain-Inspired Framework for Vision-Language-Action Models in Embodied Agents",
      "authors": "Zhejian Yang, Yongchao Chen, Xueyang Zhou, Jiangyue Yan, Dingjie Song, Yinuo Liu, Yuting Li, Yu Zhang, Pan Zhou",
      "date": "2025-05-29",
      "category": [
        "agentic-vla",
        "agent-as-policy"
      ],
      "identifier": {
        "scheme": "arXiv",
        "value": "2505.23450"
      },
      "url": "https://arxiv.org/abs/2505.23450",
      "contribution": "A coordination protocol (Standardized Action Procedure) for vision-language-action components, aimed at error accumulation and missing verification in long-horizon manipulation.",
      "relation_to_tgl": "Shares the diagnosis that execution needs verification. TGL makes the effect check part of each Skill Block's contract rather than a layer around the policy.",
      "page": "https://tgl.changnie.top/concepts/agent-as-policy/"
    },
    {
      "name": "Code as Policies: Language Model Programs for Embodied Control",
      "authors": "Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, Andy Zeng",
      "date": "2022-09-16",
      "category": [
        "coding-agent-robotics",
        "frontier-model-robot-control"
      ],
      "identifier": {
        "scheme": "arXiv",
        "value": "2209.07753"
      },
      "url": "https://arxiv.org/abs/2209.07753",
      "project_url": "https://code-as-policies.github.io/",
      "contribution": "The program-as-policy predecessor: a language model writes policy code over supplied perception primitives.",
      "relation_to_tgl": "The lineage TGL's Codex role descends from; TGL adds validation and persistence around the written program.",
      "page": "https://tgl.changnie.top/concepts/coding-agent-robotics/"
    },
    {
      "name": "Do As I Can, Not As I Say: Grounding Language in Robotic Affordances",
      "authors": "Michael Ahn et al.",
      "date": "2022-04-04",
      "category": [
        "frontier-model-robot-control"
      ],
      "identifier": {
        "scheme": "arXiv",
        "value": "2204.01691"
      },
      "url": "https://arxiv.org/abs/2204.01691",
      "project_url": "https://say-can.github.io/",
      "contribution": "Grounds language-model plans in the affordances of the robot that will execute them.",
      "relation_to_tgl": "The planning-side predecessor; TGL's grounding happens per subgoal against the current scene.",
      "page": "https://tgl.changnie.top/concepts/agent-as-policy/"
    },
    {
      "name": "ReKep: Spatio-Temporal Reasoning of Relational Keypoint Constraints for Robotic Manipulation",
      "authors": "Wenlong Huang, Chen Wang, Yunzhu Li, Ruohan Zhang, Li Fei-Fei",
      "date": "2024-09-03",
      "category": [
        "frontier-model-robot-control",
        "agent-as-policy"
      ],
      "identifier": {
        "scheme": "arXiv",
        "value": "2409.01652"
      },
      "url": "https://arxiv.org/abs/2409.01652",
      "contribution": "Represents manipulation as relational keypoint constraints and solves them in a closed loop.",
      "relation_to_tgl": "A different route to closed-loop reliability: constraints rather than validated reusable blocks.",
      "page": "https://tgl.changnie.top/concepts/agent-as-policy/"
    }
  ],
  "excluded": "Several widely-circulated 2026 items are omitted because they could not be resolved to a verifiable primary source at the time of writing. They are not listed speculatively."
}