{
  "_comment": "All numbers sourced from the ASPIRE paper (NeurIPS 2026). Tables 1-3, 7, 9-10, 13-18; Figures 6, 8, 9.",
  "liberoPro": {
    "_source": "Table 1 (page 14) for baselines and ASPIRE (Opus 4.6); Figure 9 / Tables 13-15 for ASPIRE with GPT-6 Astra and Claude Opus 5.5 (with evolutionary search, rounded). Success rates in [0, 100].",
    "suites": [
      "Object",
      "Goal",
      "Spatial"
    ],
    "axes": [
      "Pos",
      "Task"
    ],
    "methods": [
      {
        "id": "openvla",
        "label": "OpenVLA",
        "color": "#cfcfcf"
      },
      {
        "id": "pi0",
        "label": "π0",
        "color": "#aab4c2"
      },
      {
        "id": "pi05",
        "label": "π0.5",
        "color": "#93a0b2"
      },
      {
        "id": "cap0",
        "label": "CaP-Agent0",
        "color": "#6b7686"
      },
      {
        "id": "aspire",
        "label": "ASPIRE (Opus 4.6)",
        "color": "#b6df74",
        "highlight": true
      },
      {
        "id": "aspire_astra",
        "label": "ASPIRE (GPT-6 Astra)",
        "color": "#76b900",
        "highlight": true
      },
      {
        "id": "aspire_opus55",
        "label": "ASPIRE (Opus 5.5)",
        "color": "#3f7a00",
        "highlight": true
      }
    ],
    "pos": {
      "openvla": [
        0,
        0,
        0
      ],
      "pi0": [
        0,
        0,
        0
      ],
      "pi05": [
        17,
        38,
        20
      ],
      "cap0": [
        22,
        26,
        12
      ],
      "aspire": [
        98,
        81,
        51
      ],
      "aspire_astra": [
        99,
        97,
        99
      ],
      "aspire_opus55": [
        98,
        98,
        99
      ]
    },
    "task": {
      "openvla": [
        0,
        0,
        0
      ],
      "pi0": [
        0,
        0,
        0
      ],
      "pi05": [
        1,
        0,
        1
      ],
      "cap0": [
        18,
        17,
        14
      ],
      "aspire": [
        95,
        45,
        60
      ],
      "aspire_astra": [
        99,
        97,
        98
      ],
      "aspire_opus55": [
        100,
        99,
        99
      ]
    }
  },
  "robosuite": {
    "_source": "Table 2 (page 15) for Human, CaP-Agent0, ASPIRE (Opus 4.6); Table 17 (page 34) for ASPIRE with GPT-6 Astra and Claude Opus 5.5. Success rates in [0, 100] over 100 held-out trials per task.",
    "tasks": [
      "Cube Lift",
      "Cube Stack",
      "Cube Restack",
      "Spill Wipe",
      "Two-arm Handover",
      "Two-arm Lift",
      "Nut Assembly"
    ],
    "methods": [
      {
        "id": "human",
        "label": "Human",
        "color": "#94a3b8"
      },
      {
        "id": "cap0",
        "label": "CaP-Agent0",
        "color": "#6b7686"
      },
      {
        "id": "aspire",
        "label": "ASPIRE (Opus 4.6)",
        "color": "#b6df74",
        "highlight": true
      },
      {
        "id": "aspire_astra",
        "label": "ASPIRE (GPT-6 Astra)",
        "color": "#76b900",
        "highlight": true
      },
      {
        "id": "aspire_opus55",
        "label": "ASPIRE (Opus 5.5)",
        "color": "#3f7a00",
        "highlight": true
      }
    ],
    "values": {
      "human": [
        93,
        73,
        47,
        100,
        91,
        94,
        80
      ],
      "cap0": [
        97,
        98,
        89,
        100,
        20,
        74,
        0
      ],
      "aspire": [
        97,
        99,
        100,
        99,
        92,
        71,
        9
      ],
      "aspire_astra": [
        100,
        99,
        100,
        99,
        95,
        100,
        99
      ],
      "aspire_opus55": [
        100,
        100,
        100,
        100,
        100,
        100,
        100
      ]
    }
  },
  "behavior": {
    "_source": "Table 3 (page 14). Navigation and Task success rates in [0, 100] on held-out seeds. Claude Opus 5.5 was not evaluated on BEHAVIOR-1K.",
    "tasks": [
      "Soda Can pick-up",
      "Radio pick-up"
    ],
    "metrics": [
      "Nav",
      "Task"
    ],
    "series": [
      {
        "id": "human",
        "label": "Human",
        "color": "#94a3b8"
      },
      {
        "id": "cap0",
        "label": "CaP-Agent0",
        "color": "#6b7686"
      },
      {
        "id": "aspire",
        "label": "ASPIRE (Opus 4.6)",
        "color": "#b6df74",
        "highlight": true
      },
      {
        "id": "aspire_astra",
        "label": "ASPIRE (GPT-6 Astra)",
        "color": "#76b900",
        "highlight": true
      }
    ],
    "nav": {
      "human": [
        80,
        88
      ],
      "cap0": [
        84,
        80
      ],
      "aspire": [
        92,
        100
      ],
      "aspire_astra": [
        96,
        96
      ]
    },
    "task": {
      "human": [
        72,
        36
      ],
      "cap0": [
        72,
        56
      ],
      "aspire": [
        88,
        88
      ],
      "aspire_astra": [
        96,
        96
      ]
    }
  },
  "zeroshot": {
    "_source": "Table 7 (page 16). LIBERO-PRO Long zero-shot transfer; success rates in [0, 100].",
    "librarySizes": [
      0,
      25,
      50,
      90
    ],
    "aspire": {
      "pos": [
        0.0,
        5.6,
        13.8,
        22.6
      ],
      "task": [
        9.4,
        21.8,
        29.2,
        38.3
      ]
    },
    "baselines": [
      {
        "label": "OpenVLA",
        "pos": 0.0,
        "task": 0.0,
        "color": "#cfcfcf",
        "dash": "2,3"
      },
      {
        "label": "π0",
        "pos": 0.0,
        "task": 0.0,
        "color": "#aab4c2",
        "dash": "2,3"
      },
      {
        "label": "π0.5",
        "pos": 8.0,
        "task": 1.0,
        "color": "#93a0b2",
        "dash": "4,3"
      },
      {
        "label": "CaP-Agent0",
        "pos": 5.2,
        "task": 2.4,
        "color": "#6b7686",
        "dash": "6,3"
      }
    ]
  },
  "realTransfer": {
    "_source": "Table 1 (page 9). Real-robot cross-embodiment skill transfer: real-robot debugging with vs. without retrieving a simulation-discovered skill. Token counts measured until first successful program (or until budget exhausted); success over 20 held-out trials.",
    "tasks": [
      "Put bowl on plate",
      "Lift soda can",
      "Open/push drawer"
    ],
    "series": [
      {
        "id": "noskill",
        "label": "w/o Skills",
        "color": "#6b7686"
      },
      {
        "id": "skill",
        "label": "w/ Skills",
        "color": "#84cc3f",
        "highlight": true
      }
    ],
    "outputTokens": {
      "noskill": [
        0.05,
        0.18,
        1.33
      ],
      "skill": [
        0.04,
        0.03,
        0.36
      ]
    },
    "totalTokens": {
      "noskill": [
        8.65,
        61.94,
        334.917
      ],
      "skill": [
        5.11,
        6.58,
        81.67
      ]
    },
    "successPct": {
      "noskill": [
        100,
        65,
        0
      ],
      "skill": [
        100,
        95,
        55
      ]
    },
    "successRaw": {
      "noskill": [
        "20/20",
        "13/20",
        "0/20"
      ],
      "skill": [
        "20/20",
        "19/20",
        "11/20"
      ]
    }
  },
  "ablation": {
    "_source": "Figure 6(a,b) and Tables 9-10 (pages 18-19). Macro-averaged success on LIBERO-PRO suites (%).",
    "conditions": [
      {
        "id": "base",
        "label": "w/o Engine",
        "color": "#cfcfcf"
      },
      {
        "id": "engine",
        "label": "+ Engine",
        "color": "#b6df74"
      },
      {
        "id": "evo",
        "label": "+ Evo Search",
        "color": "#84cc3f"
      }
    ],
    "cumulative": {
      "pos": {
        "base": 20,
        "engine": 62,
        "evo": 76
      },
      "task": {
        "base": 9,
        "engine": 62,
        "evo": 67
      }
    },
    "increments": {
      "pos": {
        "base": 20,
        "engine": 42,
        "evo": 14
      },
      "task": {
        "base": 9,
        "engine": 53,
        "evo": 5
      }
    }
  },
  "evoIteration": {
    "_source": "Figure 6(c) and Table 11 (page 19). Per-task values are reproduced verbatim from Table 11; blank cells (search terminated early) are carried forward at the last reported value. Mean is the average over the 8 reported tasks at each iteration.",
    "iterations": [
      0,
      1,
      2,
      3,
      4
    ],
    "mean": [
      28.0,
      45.0,
      53.8,
      51.2,
      60.0
    ],
    "tasks": [
      {
        "label": "goal-swap: bowl → plate",
        "values": [
          62,
          60,
          60,
          18,
          86
        ]
      },
      {
        "label": "goal-swap: wine bottle → rack",
        "values": [
          40,
          76,
          74,
          74,
          74
        ]
      },
      {
        "label": "goal-swap: bowl → stove",
        "values": [
          62,
          82,
          82,
          82,
          82
        ]
      },
      {
        "label": "goal-task: push plate → stove",
        "values": [
          0,
          80,
          80,
          80,
          80
        ]
      },
      {
        "label": "spatial-swap: bowl on cabinet → plate",
        "values": [
          16,
          38,
          38,
          46,
          30
        ]
      },
      {
        "label": "spatial-swap: bowl next to cookie → plate",
        "values": [
          16,
          20,
          4,
          4,
          40
        ]
      },
      {
        "label": "spatial-task: bowl on cookie → plate",
        "values": [
          26,
          4,
          70,
          70,
          70
        ]
      },
      {
        "label": "spatial-task: bowl on ramekin → plate",
        "values": [
          2,
          0,
          22,
          36,
          18
        ]
      }
    ]
  },
  "modelCost": {
    "_source": "Figure 8 and Table 18 (pages 14, 35). Held-out success on the 10 LIBERO-Pro Goal tasks under position perturbation (50 seeds each), iterative debugging without evolutionary search, initially empty skill library. Total cost = cost per task x 10 at list API prices.",
    "families": [
      {
        "id": "claude",
        "label": "Claude",
        "color": "#d97757"
      },
      {
        "id": "gpt",
        "label": "GPT",
        "color": "#3b7dd8"
      },
      {
        "id": "kimi",
        "label": "Kimi (open weights)",
        "color": "#8a919c",
        "open": true
      },
      {
        "id": "qwen",
        "label": "Qwen (open weights)",
        "color": "#8b5fbf",
        "open": true
      }
    ],
    "models": [
      {
        "label": "Claude Opus 5.5",
        "short": "Opus 5.5",
        "family": "claude",
        "success": 99.2,
        "costPerTask": 18.29,
        "minutesPerTask": 51
      },
      {
        "label": "Claude Sonnet 5.5",
        "short": "Sonnet 5.5",
        "family": "claude",
        "success": 98.8,
        "costPerTask": 12.11,
        "minutesPerTask": 56
      },
      {
        "label": "Claude Fable 5.1",
        "short": "Fable 5.1",
        "family": "claude",
        "success": 96.6,
        "costPerTask": 42.24,
        "minutesPerTask": 68
      },
      {
        "label": "Claude Opus 5",
        "short": "Opus 5",
        "family": "claude",
        "success": 90.6,
        "costPerTask": 59.72,
        "minutesPerTask": 76
      },
      {
        "label": "Claude Haiku 5.5",
        "short": "Haiku 5.5",
        "family": "claude",
        "success": 68.0,
        "costPerTask": 3.97,
        "minutesPerTask": 54
      },
      {
        "label": "Claude Opus 4.6",
        "short": "Opus 4.6",
        "family": "claude",
        "success": 55.6,
        "costPerTask": 23.58,
        "minutesPerTask": 61
      },
      {
        "label": "GPT-6 Astra",
        "short": "GPT-6 Astra",
        "family": "gpt",
        "success": 94.4,
        "costPerTask": 35.37,
        "minutesPerTask": 31
      },
      {
        "label": "GPT-6.1 Sol",
        "short": "GPT-6.1 Sol",
        "family": "gpt",
        "success": 88.0,
        "costPerTask": 9.52,
        "minutesPerTask": 49
      },
      {
        "label": "GPT-6 Sol",
        "short": "GPT-6 Sol",
        "family": "gpt",
        "success": 59.2,
        "costPerTask": 7.64,
        "minutesPerTask": 42
      },
      {
        "label": "GPT-6 Luna",
        "short": "GPT-6 Luna",
        "family": "gpt",
        "success": 14.8,
        "costPerTask": 0.83,
        "minutesPerTask": 47
      },
      {
        "label": "Kimi K3",
        "short": "Kimi K3",
        "family": "kimi",
        "success": 78.4,
        "costPerTask": 58.19,
        "minutesPerTask": 139
      },
      {
        "label": "Qwen 3.8 Max",
        "short": "Qwen 3.8 Max",
        "family": "qwen",
        "success": 56.0,
        "costPerTask": 39.82,
        "minutesPerTask": 103
      }
    ]
  },
  "rubik": {
    "_source": "Figure 6(b) and Section 3.7 (page 11). Three sequential real-robot Solve-Rubik's-Cube episodes in one GPT-6 Astra session. Per completed face turn; full-episode amortized, including failed attempts, recovery, and context compaction.",
    "episodes": [
      {
        "label": "Episode 1",
        "outcome": "Unsolved",
        "tokensPerTurnM": 11.27,
        "minutesPerTurn": 16.97,
        "timePerTurn": "16 min 58 s",
        "turns": 3,
        "episodeTime": "50 min 54 s",
        "costPerTurn": 16.85
      },
      {
        "label": "Episode 2",
        "outcome": "Solved",
        "tokensPerTurnM": 5.56,
        "minutesPerTurn": 7.8,
        "timePerTurn": "7 min 48 s",
        "turns": 5,
        "episodeTime": "39 min",
        "costPerTurn": null
      },
      {
        "label": "Episode 3",
        "outcome": "Solved",
        "tokensPerTurnM": 2.01,
        "minutesPerTurn": 3.09,
        "timePerTurn": "3 min 05 s",
        "turns": 25,
        "episodeTime": "77 min 10 s",
        "costPerTurn": 2.89
      }
    ]
  },
  "modelAblation": {
    "_source": "Figure 9 and Tables 13-16 (pages 14, 31-33). Held-out success per LIBERO-Pro suite and perturbation (10 tasks x 50 seeds). 'base' is iterative debugging without evolutionary search; 'evo' adds evolutionary search. Opus 4.6 values are from Tables 9-10 and its Long values are without evolutionary search.",
    "suites": [
      "Object",
      "Goal",
      "Spatial",
      "Long"
    ],
    "axes": [
      "Pos",
      "Task"
    ],
    "models": [
      {
        "id": "opus46",
        "label": "Claude Opus 4.6",
        "color": "#dcefc0"
      },
      {
        "id": "opus5",
        "label": "Claude Opus 5",
        "color": "#b6df74"
      },
      {
        "id": "astra",
        "label": "GPT-6 Astra",
        "color": "#76b900"
      },
      {
        "id": "opus55",
        "label": "Claude Opus 5.5",
        "color": "#3f7a00"
      }
    ],
    "base": {
      "pos": {
        "opus46": [
          98,
          47,
          42,
          43.2
        ],
        "opus5": [
          98.6,
          97.0,
          97.8,
          81.0
        ],
        "astra": [
          99.0,
          97.0,
          97.0,
          80.8
        ],
        "opus55": [
          98.2,
          97.0,
          99.4,
          97.0
        ]
      },
      "task": {
        "opus46": [
          95,
          42,
          47,
          52.0
        ],
        "opus5": [
          100.0,
          86.8,
          98.0,
          83.4
        ],
        "astra": [
          99.4,
          96.8,
          97.6,
          95.0
        ],
        "opus55": [
          100.0,
          99.4,
          99.2,
          98.8
        ]
      }
    },
    "evo": {
      "pos": {
        "opus46": [
          98,
          81,
          51,
          43.2
        ],
        "opus5": [
          98.6,
          97.0,
          97.8,
          87.0
        ],
        "astra": [
          99.0,
          97.0,
          98.6,
          84.6
        ],
        "opus55": [
          98.2,
          98.0,
          99.4,
          97.0
        ]
      },
      "task": {
        "opus46": [
          95,
          45,
          60,
          52.0
        ],
        "opus5": [
          100.0,
          86.8,
          98.0,
          90.4
        ],
        "astra": [
          99.4,
          96.8,
          97.6,
          95.0
        ],
        "opus55": [
          100.0,
          99.4,
          99.2,
          98.8
        ]
      }
    }
  }
}