{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-30-009",
    "source_story_id": "tmp-story-gym-anything-robot-simulations",
    "edition_id": "mp-2026-10-30-morning-0112",
    "edition_url": "https://themachinepress.com/edition/2026-10-30",
    "position": 9,
    "story_type": "dispatch",
    "section": "robotics",
    "editorial_classification": "editorial",
    "headline": "Coding Agents Build the Robot Worlds—and Still Fail Every Test",
    "slug": "coding-agents-build-the-robot-worlds-and-still-fail-every-test",
    "dek": "RoboWorld contains 25 generated simulations and 25 test tasks; the best frontier models averaged 0.06 reward and solved none of them.",
    "summary": "RoboWorld contains 25 generated simulations and 25 test tasks; the best frontier models averaged 0.06 reward and solved none of them.",
    "body_text": "Gym-Anything for Robotics treats simulation construction as a coding-agent task. Its adversarial creation-and-audit loop checks simulations and tasks chosen to cover economically valuable physical work. The resulting RoboWorld set contains 25 simulations, 25 test environments and 175 training environments. Even the best evaluated frontier models failed every test task, averaging 0.06 reward. Repeated attempts, privileged state-reading teachers, self-assisted learning and code-generated demonstrations each improved results. Because the worlds are agent-built and audited rather than measurements of physical workplaces, the benchmark tests learning inside the released simulations, not real-world robot competence.",
    "why_it_matters": "RoboWorld contains 25 generated simulations and 25 test tasks; the best frontier models averaged 0.06 reward and solved none of them.",
    "limitations": [
      "Because the worlds are agent-built and audited rather than measurements of physical workplaces, the benchmark tests learning inside the released simulations, not real-world robot competence."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-30-009/coding-agents-build-the-robot-worlds-and-still-fail-every-test",
    "json_url": "https://themachinepress.com/story/mp-2026-10-30-009.json",
    "first_published_at": "2026-10-30T09:00:00.000-04:00",
    "modified_at": "2026-10-30T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-30-009-001",
        "text": "RoboWorld contains 25 generated simulations and 25 test tasks; the best frontier models averaged 0.06 reward and solved none of them.",
        "source_ids": [
          "source-2026-10-30-009"
        ],
        "qualification": "Because the worlds are agent-built and audited rather than measurements of physical workplaces, the benchmark tests learning inside the released simulations, not real-world robot competence."
      }
    ],
    "source_ids": [
      "source-2026-10-30-009"
    ],
    "tags": [
      "robotics",
      "simulation",
      "coding agents",
      "benchmarks"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-30-009",
      "title": "arXiv preprint 2610.40565, original v1",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.40565v1",
      "canonical_url": "https://arxiv.org/abs/2610.40565v1",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-28T20:00:00.000-04:00",
      "accessed_at": "2026-10-30T08:14:40.422-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-30-009-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Coding Agents Build the Robot Worlds—and Still Fail Every Test",
    "publisher": "The Machine Press",
    "published_at": "2026-10-30T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-30-009/coding-agents-build-the-robot-worlds-and-still-fail-every-test"
  }
}