{
 "model": "claude-opus-5-5",
 "source": "DexVerse campaign report.json files, read by scripts/export_search_tasks.py; videos are the recorded held-out success renders, copied unchanged",
 "arms": {
  "model_driven": "Agentic search",
  "cem": "CEM search",
  "baseline": "Baseline (policy search), not yet run"
 },
 "heldout_metric": "Adroit: frozen policy on 30 unseen starts, one run. DexVerse: 50 unseen starts, mean of two repeats. Unaccepted tasks count as 0%.",
 "rollouts_metric": "Total rollouts = parameter-search rollouts + policy-design resets.",
 "video_note": "One held-out episode per task, from Agentic search (video_arm: model_driven) unless noted: a success, or where Agentic search has no accepted policy, CEM search's accepted policy (video_arm: cem) or else a failed episode of the last submitted Agentic policy (video_outcome: failure); tasks with no recorded video show none.",
 "baseline_note": "Baseline values are null until the runs exist; all BASELINE entries are None",
 "summary": [
  {
   "group": "Adroit",
   "task": "Door",
   "model_driven": {
    "search_rollouts": 282,
    "resets": 1,
    "total_rollouts": 283,
    "heldout_percent": 100.0
   },
   "cem": {
    "search_rollouts": 13,
    "resets": 8,
    "total_rollouts": 21,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "group": "Adroit",
   "task": "Hammer",
   "model_driven": {
    "search_rollouts": 91,
    "resets": 2,
    "total_rollouts": 93,
    "heldout_percent": 93.3
   },
   "cem": {
    "search_rollouts": 1056,
    "resets": 7,
    "total_rollouts": 1063,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "group": "Adroit",
   "task": "Relocate",
   "model_driven": {
    "search_rollouts": 37,
    "resets": 0,
    "total_rollouts": 37,
    "heldout_percent": 66.7
   },
   "cem": {
    "search_rollouts": 31,
    "resets": 14,
    "total_rollouts": 45,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "group": "Primitive",
   "tasks": 7,
   "model_driven": {
    "search_rollouts": 119.9,
    "resets": 1.6,
    "total_rollouts": 121.4,
    "heldout_percent": 94.0
   },
   "cem": {
    "search_rollouts": 5.9,
    "resets": 6.7,
    "total_rollouts": 12.6,
    "heldout_percent": 95.6
   },
   "baseline": null
  },
  {
   "group": "Functional",
   "tasks": 9,
   "model_driven": {
    "search_rollouts": 108.0,
    "resets": 2.8,
    "total_rollouts": 110.8,
    "heldout_percent": 96.3
   },
   "cem": {
    "search_rollouts": 93.2,
    "resets": 16.3,
    "total_rollouts": 109.6,
    "heldout_percent": 98.6
   },
   "baseline": null
  },
  {
   "group": "Articulation",
   "tasks": 13,
   "model_driven": {
    "search_rollouts": 417.5,
    "resets": 6.2,
    "total_rollouts": 423.7,
    "heldout_percent": 81.0
   },
   "cem": {
    "search_rollouts": 192.5,
    "resets": 24.7,
    "total_rollouts": 217.2,
    "heldout_percent": 86.7
   },
   "baseline": null
  },
  {
   "group": "Bimanual",
   "tasks": 3,
   "model_driven": {
    "search_rollouts": 134.7,
    "resets": 0.3,
    "total_rollouts": 135.0,
    "heldout_percent": 95.3
   },
   "cem": {
    "search_rollouts": 0.0,
    "resets": 7.7,
    "total_rollouts": 7.7,
    "heldout_percent": 97.7
   },
   "baseline": null
  },
  {
   "group": "Multi-goal",
   "tasks": 34,
   "model_driven": {
    "search_rollouts": 71.2,
    "resets": 1.2,
    "total_rollouts": 72.4,
    "heldout_percent": 99.0
   },
   "cem": {
    "search_rollouts": 7.2,
    "resets": 5.5,
    "total_rollouts": 12.6,
    "heldout_percent": 98.9
   },
   "baseline": null
  }
 ],
 "tasks": [
  {
   "benchmark": "Adroit",
   "group": "Adroit",
   "task": "Pen",
   "video": "assets/videos/model-driven/Adroit-Pen.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    17
   ],
   "episodes": 30,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 4192,
    "resets": 32,
    "total_rollouts": 4224,
    "heldout_percent": 56.7
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 25699,
    "resets": 89,
    "total_rollouts": 25788,
    "heldout_percent": 33.3
   },
   "baseline": null
  },
  {
   "benchmark": "Adroit",
   "group": "Adroit",
   "task": "Door",
   "video": "assets/videos/model-driven/Adroit-Door.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    30
   ],
   "episodes": 30,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 282,
    "resets": 1,
    "total_rollouts": 283,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 13,
    "resets": 8,
    "total_rollouts": 21,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "Adroit",
   "group": "Adroit",
   "task": "Hammer",
   "video": "assets/videos/model-driven/Adroit-Hammer.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    28
   ],
   "episodes": 30,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 91,
    "resets": 2,
    "total_rollouts": 93,
    "heldout_percent": 93.3
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 1056,
    "resets": 7,
    "total_rollouts": 1063,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "Adroit",
   "group": "Adroit",
   "task": "Relocate",
   "video": "assets/videos/model-driven/Adroit-Relocate.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    20
   ],
   "episodes": 30,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 37,
    "resets": 0,
    "total_rollouts": 37,
    "heldout_percent": 66.7
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 31,
    "resets": 14,
    "total_rollouts": 45,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "RelocateObject",
   "video": "assets/videos/model-driven/DexVerse-RelocateObject.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    29,
    29
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 580,
    "resets": 2,
    "total_rollouts": 582,
    "heldout_percent": 58.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 28,
    "total_rollouts": 28,
    "heldout_percent": 69.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "PushButton",
   "video": "assets/videos/model-driven/DexVerse-PushButton.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 76,
    "resets": 0,
    "total_rollouts": 76,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "OpenFaucet",
   "video": "assets/videos/model-driven/DexVerse-OpenFaucet.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 59,
    "resets": 0,
    "total_rollouts": 59,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 41,
    "resets": 7,
    "total_rollouts": 48,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "TurnOnSwitch",
   "video": "assets/videos/model-driven/DexVerse-TurnOnSwitch.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 52,
    "resets": 1,
    "total_rollouts": 53,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 0,
    "total_rollouts": 0,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "PickCube",
   "video": "assets/videos/model-driven/DexVerse-PickCube.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 50,
    "resets": 0,
    "total_rollouts": 50,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "StackCube",
   "video": "assets/videos/model-driven/DexVerse-StackCube.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 22,
    "resets": 1,
    "total_rollouts": 23,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "PickUpStick",
   "video": "assets/videos/model-driven/DexVerse-PickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 7,
    "total_rollouts": 7,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 4,
    "total_rollouts": 4,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Primitive",
   "task": "RelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-RelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 429,
    "resets": 0,
    "total_rollouts": 429,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "GraspBleach",
   "video": "assets/videos/model-driven/DexVerse-GraspBleach.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    48
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 368,
    "resets": 16,
    "total_rollouts": 384,
    "heldout_percent": 98.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 723,
    "resets": 62,
    "total_rollouts": 785,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "FunctionalPourMug",
   "video": "assets/videos/model-driven/DexVerse-FunctionalPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 176,
    "resets": 0,
    "total_rollouts": 176,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 9,
    "total_rollouts": 9,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "FunctionalDrillApply",
   "video": "assets/videos/model-driven/DexVerse-FunctionalDrillApply.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 90,
    "resets": 0,
    "total_rollouts": 90,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 12,
    "total_rollouts": 12,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "LiftBucket",
   "video": "assets/videos/model-driven/DexVerse-LiftBucket.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 88,
    "resets": 0,
    "total_rollouts": 88,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 9,
    "total_rollouts": 9,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "GraspPan",
   "video": "assets/videos/model-driven/DexVerse-GraspPan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 80,
    "resets": 0,
    "total_rollouts": 80,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 5,
    "total_rollouts": 5,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "RemoveCupFromRack",
   "video": "assets/videos/model-driven/DexVerse-RemoveCupFromRack.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 61,
    "resets": 0,
    "total_rollouts": 61,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 91,
    "resets": 25,
    "total_rollouts": 116,
    "heldout_percent": 87.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "GraspCup",
   "video": "assets/videos/model-driven/DexVerse-GraspCup.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 50,
    "resets": 7,
    "total_rollouts": 57,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 4,
    "total_rollouts": 4,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "GraspKettle",
   "video": "assets/videos/model-driven/DexVerse-GraspKettle.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 40,
    "resets": 0,
    "total_rollouts": 40,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "FunctionalPourCan",
   "video": "assets/videos/model-driven/DexVerse-FunctionalPourCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    35,
    34
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 19,
    "resets": 2,
    "total_rollouts": 21,
    "heldout_percent": 69.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 25,
    "resets": 18,
    "total_rollouts": 43,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Functional",
   "task": "FunctionalHammerStrike",
   "video": "assets/videos/model-driven/DexVerse-FunctionalHammerStrike.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 285,
    "resets": 14,
    "total_rollouts": 299,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 6,
    "total_rollouts": 6,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "UnscrewCap",
   "video": "assets/videos/model-driven/DexVerse-UnscrewCap.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    40,
    38
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 1144,
    "resets": 10,
    "total_rollouts": 1154,
    "heldout_percent": 78.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 8,
    "total_rollouts": 8,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "OpenFlatFolder",
   "video": "assets/videos/model-driven/DexVerse-OpenFlatFolder.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    46,
    42
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 1055,
    "resets": 22,
    "total_rollouts": 1077,
    "heldout_percent": 88.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 403,
    "resets": 71,
    "total_rollouts": 474,
    "heldout_percent": 99.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "SqueezeScissors",
   "video": "assets/videos/model-driven/DexVerse-SqueezeScissors.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    49
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 962,
    "resets": 4,
    "total_rollouts": 966,
    "heldout_percent": 99.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 1689,
    "resets": 67,
    "total_rollouts": 1756,
    "heldout_percent": 68.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "LiftBasketHandle",
   "video": "assets/videos/model-driven/DexVerse-LiftBasketHandle.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    47,
    48
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 860,
    "resets": 13,
    "total_rollouts": 873,
    "heldout_percent": 95.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 14,
    "total_rollouts": 14,
    "heldout_percent": 88.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "OpenHuaweiPhone",
   "video": "assets/videos/model-driven/DexVerse-OpenHuaweiPhone.mp4",
   "video_outcome": "failure",
   "video_arm": "model_driven",
   "heldout_repeats": null,
   "episodes": 50,
   "model_driven": {
    "accepted": false,
    "search_rollouts": 594,
    "resets": 12,
    "total_rollouts": 606,
    "heldout_percent": 0.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 24,
    "total_rollouts": 24,
    "heldout_percent": 74.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "LiftLid",
   "video": "assets/videos/model-driven/DexVerse-LiftLid.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 168,
    "resets": 0,
    "total_rollouts": 168,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 7,
    "total_rollouts": 7,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "OpenGlasses",
   "video": "assets/videos/model-driven/DexVerse-OpenGlasses.mp4",
   "video_outcome": "failure",
   "video_arm": "model_driven",
   "heldout_repeats": null,
   "episodes": 50,
   "model_driven": {
    "accepted": false,
    "search_rollouts": 130,
    "resets": 17,
    "total_rollouts": 147,
    "heldout_percent": 0.0
   },
   "cem": {
    "accepted": false,
    "search_rollouts": 280,
    "resets": 74,
    "total_rollouts": 354,
    "heldout_percent": 0.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "OpenLaptop",
   "video": "assets/videos/model-driven/DexVerse-OpenLaptop.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 139,
    "resets": 0,
    "total_rollouts": 139,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 76,
    "resets": 26,
    "total_rollouts": 102,
    "heldout_percent": 98.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "OpenMicrowave",
   "video": "assets/videos/model-driven/DexVerse-OpenMicrowave.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 102,
    "resets": 0,
    "total_rollouts": 102,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 6,
    "total_rollouts": 6,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "GraspPot",
   "video": "assets/videos/model-driven/DexVerse-GraspPot.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 95,
    "resets": 0,
    "total_rollouts": 95,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "SlideUtilityKnife",
   "video": "assets/videos/model-driven/DexVerse-SlideUtilityKnife.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    46,
    47
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 92,
    "resets": 0,
    "total_rollouts": 92,
    "heldout_percent": 93.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 54,
    "resets": 19,
    "total_rollouts": 73,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "GraspBucket",
   "video": "assets/videos/model-driven/DexVerse-GraspBucket.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 86,
    "resets": 0,
    "total_rollouts": 86,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 1,
    "total_rollouts": 1,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "RotateKnob",
   "video": null,
   "video_outcome": null,
   "video_arm": null,
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 1,
    "total_rollouts": 1,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Articulation",
   "task": "OpenDrawer",
   "video": "assets/videos/model-driven/DexVerse-OpenDrawer.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 137,
    "resets": 0,
    "total_rollouts": 137,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Bimanual",
   "task": "BimanualLiftBasket",
   "video": "assets/videos/model-driven/DexVerse-BimanualLiftBasket.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    43,
    44
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 223,
    "resets": 1,
    "total_rollouts": 224,
    "heldout_percent": 87.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 6,
    "total_rollouts": 6,
    "heldout_percent": 98.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Bimanual",
   "task": "BimanualLiftDutchOven",
   "video": "assets/videos/model-driven/DexVerse-BimanualLiftDutchOven.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    49,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 97,
    "resets": 0,
    "total_rollouts": 97,
    "heldout_percent": 99.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 14,
    "total_rollouts": 14,
    "heldout_percent": 95.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Bimanual",
   "task": "BimanualLiftCarton",
   "video": "assets/videos/model-driven/DexVerse-BimanualLiftCarton.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 84,
    "resets": 0,
    "total_rollouts": 84,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Bimanual",
   "task": "BimanualLiftTray",
   "video": "assets/videos/model-driven/DexVerse-BimanualLiftTray.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    46,
    46
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 88,
    "resets": 0,
    "total_rollouts": 88,
    "heldout_percent": 92.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 8,
    "total_rollouts": 8,
    "heldout_percent": 98.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PourMugPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-PourMugPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    45,
    46
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 182,
    "resets": 1,
    "total_rollouts": 183,
    "heldout_percent": 91.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 7,
    "total_rollouts": 7,
    "heldout_percent": 95.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenDrawerPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-OpenDrawerPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 130,
    "resets": 0,
    "total_rollouts": 130,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "LiftLidPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-LiftLidPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 127,
    "resets": 0,
    "total_rollouts": 127,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 19,
    "total_rollouts": 19,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PourMugPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-PourMugPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    39,
    39
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 117,
    "resets": 5,
    "total_rollouts": 122,
    "heldout_percent": 78.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 12,
    "total_rollouts": 12,
    "heldout_percent": 89.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspBucketPourMug",
   "video": "assets/videos/model-driven/DexVerse-GraspBucketPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 121,
    "resets": 0,
    "total_rollouts": 121,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 49,
    "resets": 4,
    "total_rollouts": 53,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspPotRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-GraspPotRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 120,
    "resets": 0,
    "total_rollouts": 120,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 25,
    "resets": 7,
    "total_rollouts": 32,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspBucketPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-GraspBucketPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 119,
    "resets": 0,
    "total_rollouts": 119,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "LiftLidPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-LiftLidPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 117,
    "resets": 0,
    "total_rollouts": 117,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 11,
    "total_rollouts": 11,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "RotateKnobRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-RotateKnobRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 114,
    "resets": 1,
    "total_rollouts": 115,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 98.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "LiftLidPourMug",
   "video": "assets/videos/model-driven/DexVerse-LiftLidPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 107,
    "resets": 0,
    "total_rollouts": 107,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 7,
    "total_rollouts": 7,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenDrawerPourMug",
   "video": "assets/videos/model-driven/DexVerse-OpenDrawerPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 107,
    "resets": 0,
    "total_rollouts": 107,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 4,
    "total_rollouts": 4,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PushButtonRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-PushButtonRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 89,
    "resets": 0,
    "total_rollouts": 89,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 0,
    "total_rollouts": 0,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspPotPourMug",
   "video": "assets/videos/model-driven/DexVerse-GraspPotPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 82,
    "resets": 0,
    "total_rollouts": 82,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenMicrowavePourMug",
   "video": "assets/videos/model-driven/DexVerse-OpenMicrowavePourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 79,
    "resets": 1,
    "total_rollouts": 80,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "RotateKnobPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-RotateKnobPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 76,
    "resets": 0,
    "total_rollouts": 76,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 1,
    "total_rollouts": 1,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "LiftLidRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-LiftLidRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 74,
    "resets": 0,
    "total_rollouts": 74,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspPotPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-GraspPotPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 72,
    "resets": 0,
    "total_rollouts": 72,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PourMugRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-PourMugRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    48,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 66,
    "resets": 6,
    "total_rollouts": 72,
    "heldout_percent": 98.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 13,
    "total_rollouts": 13,
    "heldout_percent": 98.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PushButtonPourMug",
   "video": "assets/videos/model-driven/DexVerse-PushButtonPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 63,
    "resets": 0,
    "total_rollouts": 63,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenMicrowavePickUpStick",
   "video": "assets/videos/model-driven/DexVerse-OpenMicrowavePickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 61,
    "resets": 0,
    "total_rollouts": 61,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 4,
    "total_rollouts": 4,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PushButtonPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-PushButtonPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 61,
    "resets": 0,
    "total_rollouts": 61,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "RotateKnobPourMug",
   "video": "assets/videos/model-driven/DexVerse-RotateKnobPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 52,
    "resets": 1,
    "total_rollouts": 53,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 10,
    "total_rollouts": 10,
    "heldout_percent": 97.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "TurnOnSwitchPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-TurnOnSwitchPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 49,
    "resets": 4,
    "total_rollouts": 53,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenDrawerRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-OpenDrawerRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 50,
    "resets": 1,
    "total_rollouts": 51,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "RotateKnobPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-RotateKnobPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 43,
    "resets": 1,
    "total_rollouts": 44,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 11,
    "total_rollouts": 11,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "TurnOnSwitchPourMug",
   "video": "assets/videos/model-driven/DexVerse-TurnOnSwitchPourMug.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 41,
    "resets": 0,
    "total_rollouts": 41,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 9,
    "total_rollouts": 9,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "TurnOnSwitchRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-TurnOnSwitchRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 39,
    "resets": 0,
    "total_rollouts": 39,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 3,
    "total_rollouts": 3,
    "heldout_percent": 92.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenMicrowaveRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-OpenMicrowaveRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    49,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 32,
    "resets": 0,
    "total_rollouts": 32,
    "heldout_percent": 99.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 10,
    "total_rollouts": 10,
    "heldout_percent": 95.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "PushButtonPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-PushButtonPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 32,
    "resets": 0,
    "total_rollouts": 32,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 1,
    "total_rollouts": 1,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspPotPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-GraspPotPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 6,
    "total_rollouts": 6,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 5,
    "total_rollouts": 5,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenDrawerPickUpStick",
   "video": "assets/videos/model-driven/DexVerse-OpenDrawerPickUpStick.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 6,
    "total_rollouts": 6,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "OpenMicrowavePickUpCan",
   "video": "assets/videos/model-driven/DexVerse-OpenMicrowavePickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    49
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 5,
    "total_rollouts": 5,
    "heldout_percent": 99.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 170,
    "resets": 10,
    "total_rollouts": 180,
    "heldout_percent": 98.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "TurnOnSwitchPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-TurnOnSwitchPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 4,
    "total_rollouts": 4,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspBucketRelocateSphere",
   "video": "assets/videos/model-driven/DexVerse-GraspBucketRelocateSphere.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 1,
    "total_rollouts": 1,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 5,
    "total_rollouts": 5,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Multi-goal",
   "task": "GraspBucketPickUpCan",
   "video": "assets/videos/model-driven/DexVerse-GraspBucketPickUpCan.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    50,
    50
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 80,
    "resets": 0,
    "total_rollouts": 80,
    "heldout_percent": 100.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 2,
    "total_rollouts": 2,
    "heldout_percent": 100.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Long-horizon",
   "task": "LongHorizon-OvenBakeSalmon",
   "video": "assets/videos/model-driven/DexVerse-LongHorizon-OvenBakeSalmon.mp4",
   "video_outcome": "failure",
   "video_arm": "model_driven",
   "heldout_repeats": null,
   "episodes": 50,
   "model_driven": {
    "accepted": false,
    "search_rollouts": 827,
    "resets": 17,
    "total_rollouts": 844,
    "heldout_percent": 0.0
   },
   "cem": {
    "accepted": false,
    "search_rollouts": 136,
    "resets": 104,
    "total_rollouts": 240,
    "heldout_percent": 0.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Long-horizon",
   "task": "LongHorizon-MakeCoffee",
   "video": "assets/videos/model-driven/DexVerse-LongHorizon-MakeCoffee.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    44,
    44
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 78,
    "resets": 36,
    "total_rollouts": 114,
    "heldout_percent": 88.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 1214,
    "resets": 27,
    "total_rollouts": 1241,
    "heldout_percent": 90.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Long-horizon",
   "task": "LongHorizon-MicrowaveRetrievePlace",
   "video": "assets/videos/model-driven/DexVerse-LongHorizon-MicrowaveRetrievePlace.mp4",
   "video_outcome": "success",
   "video_arm": "model_driven",
   "heldout_repeats": [
    37,
    34
   ],
   "episodes": 50,
   "model_driven": {
    "accepted": true,
    "search_rollouts": 1802,
    "resets": 2,
    "total_rollouts": 1804,
    "heldout_percent": 71.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 0,
    "resets": 83,
    "total_rollouts": 83,
    "heldout_percent": 79.0
   },
   "baseline": null
  },
  {
   "benchmark": "DexVerse",
   "group": "Long-horizon",
   "task": "LongHorizon-CookFoods",
   "video": "assets/videos/model-driven/DexVerse-LongHorizon-CookFoods.mp4",
   "video_outcome": "failure",
   "video_arm": "cem",
   "heldout_repeats": null,
   "episodes": 50,
   "model_driven": {
    "accepted": false,
    "search_rollouts": 908,
    "resets": 52,
    "total_rollouts": 960,
    "heldout_percent": 0.0
   },
   "cem": {
    "accepted": true,
    "search_rollouts": 906,
    "resets": 98,
    "total_rollouts": 1004,
    "heldout_percent": 84.0
   },
   "baseline": null
  }
 ]
}
