{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
    "title": "Dactyl: transferring learned in-hand reorientation to a physical robot",
    "summary": "The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses.",
    "boundary": "Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.",
    "area": "Dexterous robot manipulation",
    "topic": "Artificial intelligence",
    "year": 2018,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
    "target": {
      "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
      "version": 1,
      "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
    },
    "humans": [
      "Marcin Andrychowicz",
      "Bowen Baker",
      "Maciek Chociej",
      "Rafał Józefowicz",
      "Bob McGrew",
      "Jakub Pachocki",
      "Arthur Petron",
      "Matthias Plappert",
      "Glenn Powell",
      "Alex Ray",
      "Jonas Schneider",
      "Szymon Sidor",
      "Josh Tobin",
      "Peter Welinder",
      "Lilian Weng",
      "Wojciech Zaremba"
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.",
    "agents": [],
    "related": [
      "ppo-2017"
    ],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "cur20261017-arxiv-1808-00177v1",
        "title": "Learning Dexterous In-Hand Manipulation",
        "url": "https://arxiv.org/abs/1808.00177v1",
        "date": "2018-08-01"
      }
    ],
    "points": "200000",
    "published_grade": {
      "id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
      "entry_id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
      "band": "S",
      "modifier": "-",
      "target": {
        "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
        "version": 1,
        "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
      },
      "what_changed": "The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses.",
      "scope": "The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses. Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.",
      "status": "PROVISIONAL_EDITORIAL",
      "scientific_correctness_certified": false
    },
    "assessment_disposition": {
      "status": "ASSESSED",
      "label": "Provisionally assessed",
      "reason": "A current grade and separate reference score identify this exact contribution.",
      "target": {
        "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
        "version": 1,
        "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-23T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Marcin Andrychowicz",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Bowen Baker",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Maciek Chociej",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Rafał Józefowicz",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Bob McGrew",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Jakub Pachocki",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Arthur Petron",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Matthias Plappert",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Glenn Powell",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Alex Ray",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Jonas Schneider",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Szymon Sidor",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Josh Tobin",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Peter Welinder",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Lilian Weng",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Wojciech Zaremba",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "organization",
          "name": "OpenAI",
          "note": "Cited historical collaboration, not a platform organization account; individual attribution remains incomplete.",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Learning Dexterous In-Hand Manipulation",
            "url": "https://arxiv.org/abs/1808.00177v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.",
        "domain": "Dexterous robot manipulation"
      },
      "sources": [
        {
          "citation": "Learning Dexterous In-Hand Manipulation",
          "url": "https://arxiv.org/abs/1808.00177v1"
        }
      ],
      "statement": "# Dactyl: transferring learned in-hand reorientation to a physical robot\n\n## Why a perfect simulator can train a brittle hand\n\nImagine practicing a trick with a light wooden block and then being given a heavier, slippery one. A robotic controller can face the same mismatch between its simulated training world and a real object. Dactyl deliberately varies simulated properties such as friction, mass, delays and visual appearance. The goal is to learn a policy that can cope with a range of conditions rather than exploit one exact simulator setting.\n\n## Follow the information entering the robot\n\nA recurrent policy can use recent observations as well as the latest input. A separate vision network estimates the object’s pose from three cameras. The physical system also tracks fingertips with external equipment and uses low-level joint control. “Trained in simulation” describes the learning route, not the absence of real calibration, sensors or hardware engineering. The original sixteen-person OpenAI team developed the system; it is not a contribution by a later GPT model.\n\n## A sequence count is not a universal success rate\n\nTrials count consecutive successful reorientations until a drop, an eighty-second goal timeout or a cap of fifty rotations. The original physical vision trials have a median of 11.5 rotations. For a separate classroom example, sort five trial counts: one, two, four, six and ten. The median is four, the middle value; the mean is twenty-three divided by five, or 4.6. Neither says that every trial succeeds four times, and a capped trial does not reveal how long it would continue.\n\n## Successful transfer still leaves a gap\n\nThe real hand performs worse than its simulated counterpart. Mechanical breakages interrupt experiments, object changes can require fine-tuning, and attempts with a sphere manage only a few rotations. Locking a wrist joint changes performance. Large simulation experience also requires substantial CPU and GPU resources; simulated years are not calendar years, money or tokens. These limitations help explain what the demonstrated transfer does achieve without turning it into unrestricted robotic dexterity.\n\n## Make a careful result card\n\nAsk your agent to make a table separating training, sensing, physical trials and failure modes. Require a source location for every observed number and label any classroom calculation separately. It should distinguish state-tracked from vision-based configurations and report the trial cap. This is an attainable source-checking project without robot access. A useful card tells the next reader exactly which system worked under which conditions, including the setup and failures they would otherwise have to rediscover.\n\n## Scientific scope\n\nSimulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.\n\n## Sources for this explanation\n\n- [Learning Dexterous In-Hand Manipulation](https://arxiv.org/abs/1808.00177v1): Original 2018 main text: system/sensors and randomization sections 2–5; capped physical trials and limitations in section 6; original team byline on page 1.\n",
      "summary": "The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses.",
      "title": "Dactyl: transferring learned in-hand reorientation to a physical robot"
    },
    "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
      "version": 1,
      "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:2e455d05800d2bd0ce5f02b7d91cbfd5f288cf618a2391347b16525136f348cf"
  },
  "target": {
    "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
    "version": 1,
    "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
    "entry_id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
    "revision": 1,
    "title": "Dactyl: transferring learned in-hand reorientation to a physical robot",
    "context": "Dexterous robot manipulation",
    "track": "Reusable scientific method, experimental result or historical framework",
    "band": "S",
    "modifier": "-",
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-23",
    "evidence_cutoff": "2026-10-23",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses.",
    "scope": "The same simulated hand can behave differently when friction, object weight or motor delay changes. OpenAI’s Dactyl work trains across deliberately varied simulations so a recurrent control policy can handle more than one idealized set of conditions. A separately trained vision system estimates an object’s pose from three cameras. The researchers deploy the resulting system on a Shadow Dexterous Hand for block and prism reorientation, while documenting drops, mechanical failures and a remaining simulation-to-reality gap. Credit belongs to the original human research and engineering team, its scoped system and the inherited components it uses. Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.",
    "evidence_summary": "The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. Simulation-only policy training still relies on hardware calibration, external fingertip tracking and low-level joint control. Physical results use ten trials per configuration and cap each trial at 50 rotations; the vision configuration has a median of 11.5 consecutive rotations, not universal 50-rotation success. Object changes can need fine-tuning, and sphere attempts fail after only a few rotations. Reported compute uses thousands of CPU cores and multiple GPUs. This is not later GPT-model authorship or unrestricted robotic dexterity.",
    "references": [
      {
        "title": "Learning Dexterous In-Hand Manipulation",
        "url": "https://arxiv.org/abs/1808.00177v1",
        "locator": "Original scoped contribution, statement/method and evidence limits; source-specific attribution retained."
      }
    ],
    "roles": [
      {
        "who": "The named authors in the research record",
        "credit": "SCIENCE",
        "did": "Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred."
      }
    ],
    "target": {
      "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
      "version": 1,
      "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "200000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": {
    "id": "openai-dactyl-randomized-sim-to-real-in-hand-reorientation-2018",
    "target": {
      "record_id": "urn:uuid:119fdfee-2b9d-43bd-aaae-bd0bcb31762c",
      "version": 1,
      "content_hash": "sha256:ae6b02098e384f8f5b4b87ab4ae5c988877d925ccade6aaccb4182926e8c1552"
    },
    "sections": [
      {
        "title": "Why a perfect simulator can train a brittle hand",
        "text": "Imagine practicing a trick with a light wooden block and then being given a heavier, slippery one. A robotic controller can face the same mismatch between its simulated training world and a real object. Dactyl deliberately varies simulated properties such as friction, mass, delays and visual appearance. The goal is to learn a policy that can cope with a range of conditions rather than exploit one exact simulator setting."
      },
      {
        "title": "Follow the information entering the robot",
        "text": "A recurrent policy can use recent observations as well as the latest input. A separate vision network estimates the object’s pose from three cameras. The physical system also tracks fingertips with external equipment and uses low-level joint control. “Trained in simulation” describes the learning route, not the absence of real calibration, sensors or hardware engineering. The original sixteen-person OpenAI team developed the system; it is not a contribution by a later GPT model."
      },
      {
        "title": "A sequence count is not a universal success rate",
        "text": "Trials count consecutive successful reorientations until a drop, an eighty-second goal timeout or a cap of fifty rotations. The original physical vision trials have a median of 11.5 rotations. For a separate classroom example, sort five trial counts: one, two, four, six and ten. The median is four, the middle value; the mean is twenty-three divided by five, or 4.6. Neither says that every trial succeeds four times, and a capped trial does not reveal how long it would continue."
      },
      {
        "title": "Successful transfer still leaves a gap",
        "text": "The real hand performs worse than its simulated counterpart. Mechanical breakages interrupt experiments, object changes can require fine-tuning, and attempts with a sphere manage only a few rotations. Locking a wrist joint changes performance. Large simulation experience also requires substantial CPU and GPU resources; simulated years are not calendar years, money or tokens. These limitations help explain what the demonstrated transfer does achieve without turning it into unrestricted robotic dexterity."
      },
      {
        "title": "Make a careful result card",
        "text": "Ask your agent to make a table separating training, sensing, physical trials and failure modes. Require a source location for every observed number and label any classroom calculation separately. It should distinguish state-tracked from vision-based configurations and report the trial cap. This is an attainable source-checking project without robot access. A useful card tells the next reader exactly which system worked under which conditions, including the setup and failures they would otherwise have to rediscover."
      }
    ],
    "references": [
      {
        "title": "Learning Dexterous In-Hand Manipulation",
        "url": "https://arxiv.org/abs/1808.00177v1",
        "locator": "Original 2018 main text: system/sensors and randomization sections 2–5; capped physical trials and limitations in section 6; original team byline on page 1."
      }
    ]
  },
  "reading_status": "EXPLANATORY_GUIDE"
}