{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-15-002",
    "source_story_id": "tmp-feature-shieldvla-reachability-safety",
    "edition_id": "mp-2026-09-15-morning-0068",
    "edition_url": "https://themachinepress.com/edition/2026-09-15",
    "position": 2,
    "story_type": "secondary",
    "section": "robotics",
    "editorial_classification": "editorial",
    "headline": "The Robot Learned Where Reward Had to Stop",
    "slug": "the-robot-learned-where-reward-had-to-stop",
    "dek": "A reachability-based safety critic cut cumulative safety cost 57% on average while improving task success by 0.13 over SafeVLA.",
    "summary": "A reachability-based safety critic cut cumulative safety cost 57% on average while improving task success by 0.13 over SafeVLA.",
    "body_text": "Safety fine-tuning for vision-language-action models often treats violations as a soft penalty, forcing one objective to trade reward against risk. ShieldVLA instead learns a model-free approximation of a Hamilton–Jacobi reachability value function from visual observations. The critic separates ordinary reward optimization inside the feasible region from recovery behavior near unsafe states; rubric-based vision-language scores provide training targets without dense manual cost labels. Across five navigation and manipulation benchmarks and multiple VLA backbones, the authors report 57% lower cumulative safety cost on average and a 0.13 gain in task success over SafeVLA. Those benchmark results do not constitute a formal guarantee for an untested physical deployment.",
    "why_it_matters": "A reachability-based safety critic cut cumulative safety cost 57% on average while improving task success by 0.13 over SafeVLA.",
    "limitations": [
      "Those benchmark results do not constitute a formal guarantee for an untested physical deployment."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-15-002/the-robot-learned-where-reward-had-to-stop",
    "json_url": "https://themachinepress.com/story/mp-2026-09-15-002.json",
    "first_published_at": "2026-09-15T09:00:00.000-04:00",
    "modified_at": "2026-09-15T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-15-002-001",
        "text": "A reachability-based safety critic cut cumulative safety cost 57% on average while improving task success by 0.13 over SafeVLA.",
        "source_ids": [
          "source-2026-09-15-002"
        ],
        "qualification": "Those benchmark results do not constitute a formal guarantee for an untested physical deployment."
      }
    ],
    "source_ids": [
      "source-2026-09-15-002"
    ],
    "tags": [
      "vision-language-action",
      "robot safety",
      "reachability"
    ],
    "image_url": "https://themachinepress.com/issues/2026-09-15/feature-shieldvla.png",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-15-002",
      "title": "arXiv preprint 2609.13231",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.13231",
      "canonical_url": "https://arxiv.org/abs/2609.13231",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-13T20:00:00.000-04:00",
      "accessed_at": "2026-09-15T08:18:00.000-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-15-002-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Robot Learned Where Reward Had to Stop",
    "publisher": "The Machine Press",
    "published_at": "2026-09-15T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-15-002/the-robot-learned-where-reward-had-to-stop"
  }
}