{
  "@context": {
    "@language": "en",
    "@vocab": "https://schema.org/",
    "citeAs": "cr:citeAs",
    "column": "cr:column",
    "conformsTo": "dct:conformsTo",
    "cr": "http://mlcommons.org/croissant/",
    "rai": "http://mlcommons.org/croissant/RAI/",
    "data": {
      "@id": "cr:data",
      "@type": "@json"
    },
    "dataType": {
      "@id": "cr:dataType",
      "@type": "@vocab"
    },
    "equivalentProperty": {
      "@id": "cr:equivalentProperty",
      "@type": "@vocab"
    },
    "samplingRate": "cr:samplingRate",
    "dct": "http://purl.org/dc/terms/",
    "examples": {
      "@id": "cr:examples",
      "@type": "@json"
    },
    "extract": "cr:extract",
    "field": "cr:field",
    "fileProperty": "cr:fileProperty",
    "fileObject": "cr:fileObject",
    "fileSet": "cr:fileSet",
    "format": "cr:format",
    "includes": "cr:includes",
    "isLiveDataset": "cr:isLiveDataset",
    "jsonPath": "cr:jsonPath",
    "key": "cr:key",
    "md5": "cr:md5",
    "parentField": "cr:parentField",
    "path": "cr:path",
    "recordSet": "cr:recordSet",
    "references": "cr:references",
    "regex": "cr:regex",
    "repeated": "cr:repeated",
    "replace": "cr:replace",
    "sc": "https://schema.org/",
    "separator": "cr:separator",
    "source": "cr:source",
    "subField": "cr:subField",
    "transform": "cr:transform",
    "prov": "http://www.w3.org/ns/prov#"
  },
  "@type": "sc:Dataset",
  "name": "EmbodiedRestore",
  "description": "First-frame robotic observations rendered in Robosuite/MuJoCo and corrupted by 25 distortions from the TID2013/KADID-10k taxonomy. Each episode provides paired low-quality (LQ) and ground-truth (GT) images from base and wrist cameras, evaluated under three policies (pi0.5, pi0, openvla); per-rollout success and step counts are exposed via results.csv.",
  "conformsTo": "http://mlcommons.org/croissant/1.0",
  "url": "https://huggingface.co/datasets/qruisjtu/EmbodiedRestore",
  "license": "https://creativecommons.org/licenses/by/4.0/",
  "version": "0.2.0",
  "datePublished": "2026-05-05",
  "creator": [
    {
      "@type": "sc:Person",
      "name": "Rui Qing, Xinpeng Liu, Chunyi Li, Jianbo Zhang, Xiaogang Xu, Xingrui Yu, Ivor Tsang, Guangtao Zhai"
    }
  ],
  "citeAs": "TODO: arXiv / NeurIPS bibtex",
  "keywords": [
    "image restoration",
    "image quality assessment",
    "robot learning",
    "robustness",
    "Robosuite",
    "MuJoCo",
    "TID2013",
    "KADID-10k",
    "pi0.5",
    "pi0",
    "openvla"
  ],
  "rai:dataCollection": "Synthetic data: scenes were rendered in Robosuite (MIT) on top of MuJoCo (Apache-2.0). For every rollout we save the first frame from two cameras (base, wrist), then apply each of 25 distortions (TID2013/KADID-10k taxonomy) to produce LQ images while keeping the original render as GT. Three policies (pi0.5, pi0, openvla) were each evaluated independently, producing parallel image folders and per-distortion result.json files.",
  "rai:dataCollectionType": "Synthetic",
  "rai:dataCollectionTimeframe": [
    "2026-01-01",
    "2026-04-30"
  ],
  "rai:dataAnnotationProtocol": "No human annotation. Image labels (policy, distort_id, eps, camera, role) are derived deterministically from filenames; rollout outcomes (success / steps / scene materials / target object) come directly from the simulator and are stored verbatim in each result.json (then flattened into results.csv).",
  "rai:dataAnnotationPlatform": "N/A (programmatic)",
  "rai:dataAnnotationAnalysis": "N/A",
  "rai:dataPreprocessingProtocol": [
    "Render first-frame RGB from Robosuite scenes at the simulator's native resolution.",
    "Apply each of 25 distortions with implementations matching TID2013/KADID-10k references.",
    "Save as PNG; pair LQ/GT per (policy, episode, camera).",
    "Flatten per-policy per-distortion result.json into results.csv for tabular consumption."
  ],
  "rai:dataUseCases": "Primary: benchmarking image restoration / IQA algorithms on robot-observation distributions. Secondary: studying how restoration quality translates to downstream policy success rate across three policies (pi0.5, pi0, openvla); see results.csv for the rollout-level table and the per-distortion result.json files for the raw success_rate plus env_metadata.",
  "rai:dataLimitations": "[\"Only the first frame of each rollout is kept; no temporal information.\", \"Single simulator (Robosuite/MuJoCo); domain gap to real cameras is not characterized.\", \"Three policies (pi0.5, pi0, openvla) but a single task suite; rankings may not generalize.\", \"Distortion strength is fixed per category; severity sweeps are not provided.\"]",
  "rai:dataBiases": "[\"Scene/object distribution inherited from the underlying Robosuite tasks.\", \"Class imbalance across distortion families (noise has more categories than blur).\", \"Camera intrinsics and lighting reflect simulator defaults, not real hardware.\"]",
  "rai:dataSocialImpact": "Low direct social impact: synthetic robotic frames with no humans. Indirect benefit through more robust perception for assistive robotics.",
  "rai:personalSensitiveInformation": "None. All images are synthetic; no humans, faces, voices, or PII.",
  "rai:dataReleaseMaintenancePlan": "Hosted at the URL above. The authors will fix metadata issues reported via GitHub issues for at least 2 years post-publication. Versioned releases via tags.",
  "distribution": [
    {
      "@type": "cr:FileObject",
      "@id": "manifest.csv",
      "name": "manifest.csv",
      "description": "Per-image manifest with policy, distortion id, episode, camera, role.",
      "contentUrl": "manifest.csv",
      "encodingFormat": "text/csv",
      "sha256": "4598fd6319b97d39b9b39f3c01bd5e1ba5672b84a34084148b7d8f97ea55686a"
    },
    {
      "@type": "cr:FileObject",
      "@id": "distort_taxonomy.csv",
      "name": "distort_taxonomy.csv",
      "description": "Mapping of distort_id to name, family, reference benchmark.",
      "contentUrl": "distort_taxonomy.csv",
      "encodingFormat": "text/csv",
      "sha256": "ad04e20a07b9f6d272a86aa48db9d00bdfe72e7336b58c1f9c8c22063ec613e9"
    },
    {
      "@type": "cr:FileObject",
      "@id": "results.csv",
      "name": "results.csv",
      "description": "Rollout-level outcomes flattened from every <policy>_DistortOnly/distort_<id>/result.json. Columns: policy, distort_id, distort_name, eps, wall_material, table_material, object, success, steps, success_rate. Joins with manifest.csv on (policy, distort_id, eps).",
      "contentUrl": "results.csv",
      "encodingFormat": "text/csv",
      "sha256": "b58733153c123ea902a2aa82aa9c95e4c6fc6b2a9e15f85247d721fec9b2ae5c"
    },
    {
      "@type": "cr:FileSet",
      "@id": "policy_results",
      "name": "policy_results",
      "description": "Raw per-(policy, distortion) result.json files containing the distortion-level success_rate plus 100-element env_metadata array.",
      "encodingFormat": "application/json",
      "includes": "*_DistortOnly/distort_*/result.json"
    },
    {
      "@type": "cr:FileSet",
      "@id": "images",
      "name": "images",
      "description": "All PNG frames across 3 policies x 25 distortions x 100 episodes x 4 views.",
      "encodingFormat": "image/png",
      "includes": "*_DistortOnly/distort_*/firstframe_distort_*_eps*_*.png"
    }
  ],
  "recordSet": [
    {
      "@type": "cr:RecordSet",
      "@id": "frames",
      "name": "frames",
      "description": "One record per PNG frame.",
      "key": {
        "@id": "frames/path"
      },
      "field": [
        {
          "@type": "cr:Field",
          "@id": "frames/path",
          "name": "path",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "path"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/policy",
          "name": "policy",
          "description": "pi0.5 | pi0 | openvla",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "policy"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/distort_id",
          "name": "distort_id",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "distort_id"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/distort_name",
          "name": "distort_name",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "distort_name"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/eps",
          "name": "eps",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "eps"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/camera",
          "name": "camera",
          "description": "base | wrist",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "camera"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/role",
          "name": "role",
          "description": "lq (distorted) | gt (clean)",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "role"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "frames/image",
          "name": "image",
          "dataType": "sc:ImageObject",
          "source": {
            "fileSet": {
              "@id": "images"
            },
            "extract": {
              "fileProperty": "content"
            }
          },
          "references": {
            "fileObject": {
              "@id": "manifest.csv"
            },
            "extract": {
              "column": "path"
            }
          }
        }
      ]
    },
    {
      "@type": "cr:RecordSet",
      "@id": "rollouts",
      "name": "rollouts",
      "description": "One record per simulated rollout (3 policies x 25 distortions x 100 episodes = 7500 rows). Joins with the frames RecordSet on (policy, distort_id, eps).",
      "field": [
        {
          "@type": "cr:Field",
          "@id": "rollouts/policy",
          "name": "policy",
          "description": "pi0.5 | pi0 | openvla",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "policy"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/distort_id",
          "name": "distort_id",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "distort_id"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/distort_name",
          "name": "distort_name",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "distort_name"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/eps",
          "name": "eps",
          "description": "1..100, episode index in filename convention; env_metadata[i-1] in result.json corresponds to eps i",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "eps"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/wall_material",
          "name": "wall_material",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "wall_material"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/table_material",
          "name": "table_material",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "table_material"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/object",
          "name": "object",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "object"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/success",
          "name": "success",
          "description": "literal 'success' or 'failure' as emitted by the simulator",
          "dataType": "sc:Text",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "success"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/steps",
          "name": "steps",
          "description": "rollout length in env steps; capped at 250 on failure",
          "dataType": "sc:Integer",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "steps"
            }
          }
        },
        {
          "@type": "cr:Field",
          "@id": "rollouts/success_rate",
          "name": "success_rate",
          "description": "Distortion-level success rate for (policy, distort_id); repeated across all 100 rows of the same group.",
          "dataType": "sc:Float",
          "source": {
            "fileObject": {
              "@id": "results.csv"
            },
            "extract": {
              "column": "success_rate"
            }
          }
        }
      ]
    }
  ],
  "rai:hasSyntheticData": true,
  "prov:wasDerivedFrom": [
    {
      "@id": "https://drive.google.com/drive/folders/1uU1vnxE5rk-ok8MNgQ1sgFkZB5UWx3nf",
      "prov:label": "EmbodiedComp",
      "sc:license": "CC-BY-4.0"
    }
  ],
  "prov:wasGeneratedBy": [
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q4929239"
      },
      "prov:label": "Scene rendering",
      "sc:description": "Robotic manipulation scenes were rendered in https://github.com/ARISE-Initiative/robosuite (MIT licence) on top of the\n   https://github.com/google-deepmind/mujoco physics engine (Apache-2.0). Each scene varies the wall material, table\n  material, and target object drawn from Robosuite's built-in asset library. For every rollout we capture the first\n  frame from two simulated cameras: a base (third-person) camera and a wrist (egocentric) camera, at the simulator's\n  native resolution.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:SoftwareAgent",
          "@id": "robosuite",
          "prov:label": "robosuite",
          "sc:description": "automated rendering pipeline executed by the authors; no human-in-the-loop."
        }
      ]
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q5227332"
      },
      "prov:label": "Distortion synthesis",
      "sc:description": "For every captured first frame we apply 25 image distortions drawn from the TID2013 and KADID-10k taxonomies, covering\n   ten families: noise, blur, denoising, compression, pattern, structural, intensity_color, color, optical, and\n  sampling. Each distortion is implemented as a deterministic image-processing function with a fixed strength per\n  category (no severity sweep). The original render is retained as the ground-truth (GT) image; the distorted output is\n  the low-quality (LQ) image, paired by (policy, episode, camera). Agent: the authors' Python distortion library,\n  executed offline; no learned model is involved in producing the LQ images."
    },
    {
      "@type": "prov:Activity",
      "prov:type": {
        "@id": "https://www.wikidata.org/wiki/Q5227332"
      },
      "prov:label": "Metadata flattening",
      "sc:description": "A single Python script (build_metadata.py) walks the three policy directories and emits: manifest.csv (one row per\n  PNG, 30,000 rows), results.csv (one row per rollout, 7,500 rows, joining manifest.csv on (policy, distort_id, eps)),\n  distort_taxonomy.csv, and the Croissant 1.0 + RAI metadata document croissant.json. Labels in manifest.csv (policy,\n  distort_id, distort_name, eps, camera, role) are derived deterministically from filenames and directory layout — there\n   is no human or model-based annotation step.",
      "prov:wasAttributedTo": [
        {
          "@type": "prov:SoftwareAgent",
          "@id": "build_metadata.py",
          "prov:label": "build_metadata.py",
          "sc:description": "the authors' code, executed locally"
        }
      ]
    }
  ]
}