{
  "source": "https://arxiv.org/abs/2610.13117",
  "provenance": "Transcribed from Tables 2–5 and Figure 5; verified against the published paper on 2026-10-13. Values retain the reported precision; differences in the interactive success charts are percentage-point differences.",
  "datasets": {
    "real": {
      "label": "Real robot",
      "title": "Five tasks on a real robot",
      "source": "Figure 5",
      "href": "https://arxiv.org/pdf/2610.13117#page=8",
      "defaultMetric": "average",
      "note": "Dual-arm ARX-X5 · 20 trials per method per task · averages across five tasks.",
      "metrics": [
        {
          "key": "drawer",
          "label": "Drawer packing",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "box",
          "label": "Box packing",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "shuttlecock",
          "label": "Shuttlecock insertion",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "cup",
          "label": "Cup stacking",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "block",
          "label": "Block stacking",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "average",
          "label": "Average",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        }
      ],
      "rows": [
        {
          "name": "π₀",
          "family": "vlm",
          "values": {
            "drawer": 60,
            "box": 65,
            "shuttlecock": 55,
            "cup": 80,
            "block": 75,
            "average": 67
          },
          "detail": "VLA baseline",
          "ours": false
        },
        {
          "name": "π₀.₅",
          "family": "vlm",
          "values": {
            "drawer": 75,
            "box": 80,
            "shuttlecock": 75,
            "cup": 95,
            "block": 90,
            "average": 83
          },
          "detail": "VLA baseline",
          "ours": false
        },
        {
          "name": "FDM w/ VLM",
          "family": "vlm",
          "values": {
            "drawer": 95,
            "box": 95,
            "shuttlecock": 90,
            "cup": 100,
            "block": 100,
            "average": 96
          },
          "detail": "Our model",
          "ours": true
        },
        {
          "name": "DreamZero",
          "family": "vgm",
          "values": {
            "drawer": 65,
            "box": 70,
            "shuttlecock": 65,
            "cup": 75,
            "block": 80,
            "average": 71
          },
          "detail": "WAM baseline",
          "ours": false
        },
        {
          "name": "FastWAM-Joint",
          "family": "vgm",
          "values": {
            "drawer": 70,
            "box": 70,
            "shuttlecock": 75,
            "cup": 90,
            "block": 85,
            "average": 78
          },
          "detail": "Future-conditioned WAM",
          "ours": false
        },
        {
          "name": "FDM w/ VGM",
          "family": "vgm",
          "values": {
            "drawer": 70,
            "box": 75,
            "shuttlecock": 70,
            "cup": 100,
            "block": 100,
            "average": 83
          },
          "detail": "Our model",
          "ours": true
        }
      ]
    },
    "libero": {
      "label": "LIBERO",
      "title": "Four suites of manipulation tasks",
      "source": "Table 2",
      "href": "https://arxiv.org/pdf/2610.13117#page=6",
      "defaultMetric": "average",
      "note": "Spatial, Object, Goal and Long suites. P.T. denotes robot policy pretraining; both FDM variants use no robot policy pretraining.",
      "metrics": [
        {
          "key": "spatial",
          "label": "Spatial",
          "digits": 1,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "object",
          "label": "Object",
          "digits": 1,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "goal",
          "label": "Goal",
          "digits": 1,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "long",
          "label": "Long",
          "digits": 1,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "average",
          "label": "Average",
          "digits": 1,
          "unit": "%",
          "max": 100,
          "higher": true
        }
      ],
      "rows": [
        {
          "name": "π₀",
          "family": "vlm",
          "values": {
            "spatial": 96.8,
            "object": 98.8,
            "goal": 95.8,
            "long": 85.2,
            "average": 94.1
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "π₀.₅",
          "family": "vlm",
          "values": {
            "spatial": 98.8,
            "object": 98.2,
            "goal": 98.0,
            "long": 92.4,
            "average": 96.9
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "FDM w/ VLM",
          "family": "vlm",
          "values": {
            "spatial": 99.2,
            "object": 100,
            "goal": 99.8,
            "long": 98.6,
            "average": 99.4
          },
          "detail": "Without robot policy pretraining",
          "ours": true
        },
        {
          "name": "Motus",
          "family": "vgm",
          "values": {
            "spatial": 96.8,
            "object": 99.8,
            "goal": 96.6,
            "long": 97.6,
            "average": 97.7
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "LingBot-VA",
          "family": "vgm",
          "values": {
            "spatial": 98.5,
            "object": 99.6,
            "goal": 97.2,
            "long": 98.5,
            "average": 98.5
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "FDM w/ VGM",
          "family": "vgm",
          "values": {
            "spatial": 98.8,
            "object": 100,
            "goal": 99.2,
            "long": 94.2,
            "average": 98.1
          },
          "detail": "Without robot policy pretraining",
          "ours": true
        }
      ]
    },
    "robotwin": {
      "label": "RoboTwin 2.0",
      "title": "Bimanual manipulation across 50 tasks",
      "source": "Table 3",
      "href": "https://arxiv.org/pdf/2610.13117#page=6",
      "defaultMetric": "average",
      "note": "50 tasks · clean and randomized conditions receive equal weight in the average · both FDM variants use no robot policy pretraining.",
      "metrics": [
        {
          "key": "clean",
          "label": "Clean",
          "digits": 2,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "randomized",
          "label": "Randomized",
          "digits": 2,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "average",
          "label": "Average",
          "digits": 2,
          "unit": "%",
          "max": 100,
          "higher": true
        }
      ],
      "rows": [
        {
          "name": "π₀",
          "family": "vlm",
          "values": {
            "clean": 65.92,
            "randomized": 58.4,
            "average": 62.16
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "π₀.₅",
          "family": "vlm",
          "values": {
            "clean": 82.74,
            "randomized": 76.76,
            "average": 79.75
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "FDM w/ VLM",
          "family": "vlm",
          "values": {
            "clean": 81.54,
            "randomized": 80.86,
            "average": 81.2
          },
          "detail": "Without robot policy pretraining",
          "ours": true
        },
        {
          "name": "Motus",
          "family": "vgm",
          "values": {
            "clean": 88.66,
            "randomized": 87.02,
            "average": 87.8
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "LingBot-VA",
          "family": "vgm",
          "values": {
            "clean": 92.9,
            "randomized": 91.5,
            "average": 92.2
          },
          "detail": "With robot policy pretraining",
          "ours": false
        },
        {
          "name": "FDM w/ VGM",
          "family": "vgm",
          "values": {
            "clean": 88.84,
            "randomized": 89.62,
            "average": 89.23
          },
          "detail": "Without robot policy pretraining",
          "ours": true
        }
      ]
    },
    "ablations": {
      "label": "Ablations",
      "title": "Isolate the contribution of future gradients",
      "source": "Table 4",
      "href": "https://arxiv.org/pdf/2610.13117#page=8",
      "defaultMetric": "average",
      "note": "Matched training settings and seeds within each backbone. All variants deploy with action-only inference. Full FDM versus stop-gradient isolates backward supervision.",
      "metrics": [
        {
          "key": "drawer",
          "label": "Drawer packing",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "box",
          "label": "Box packing",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "shuttlecock",
          "label": "Shuttlecock insertion",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "cup",
          "label": "Cup stacking",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "block",
          "label": "Block stacking",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        },
        {
          "key": "average",
          "label": "Average",
          "digits": 0,
          "unit": "%",
          "max": 100,
          "higher": true
        }
      ],
      "rows": [
        {
          "name": "A. Action-only",
          "family": "vlm",
          "values": {
            "drawer": 75,
            "box": 80,
            "shuttlecock": 75,
            "cup": 95,
            "block": 90,
            "average": 83
          },
          "detail": "Action imitation only",
          "ours": false
        },
        {
          "name": "B. Context-only future",
          "family": "vlm",
          "values": {
            "drawer": 85,
            "box": 80,
            "shuttlecock": 80,
            "cup": 95,
            "block": 95,
            "average": 87
          },
          "detail": "Future objective, without action features",
          "ours": false
        },
        {
          "name": "C. Stop-gradient",
          "family": "vlm",
          "values": {
            "drawer": 80,
            "box": 85,
            "shuttlecock": 85,
            "cup": 95,
            "block": 100,
            "average": 89
          },
          "detail": "Action features, with future gradients stopped",
          "ours": false
        },
        {
          "name": "D. Full FDM",
          "family": "vlm",
          "values": {
            "drawer": 95,
            "box": 95,
            "shuttlecock": 90,
            "cup": 100,
            "block": 100,
            "average": 96
          },
          "detail": "Action features + future-loss gradients",
          "ours": true
        },
        {
          "name": "A. Action-only",
          "family": "vgm",
          "values": {
            "drawer": 50,
            "box": 60,
            "shuttlecock": 55,
            "cup": 85,
            "block": 80,
            "average": 66
          },
          "detail": "Action imitation only",
          "ours": false
        },
        {
          "name": "B. Context-only future",
          "family": "vgm",
          "values": {
            "drawer": 60,
            "box": 65,
            "shuttlecock": 55,
            "cup": 90,
            "block": 90,
            "average": 72
          },
          "detail": "Future objective, without action features",
          "ours": false
        },
        {
          "name": "C. Stop-gradient",
          "family": "vgm",
          "values": {
            "drawer": 70,
            "box": 70,
            "shuttlecock": 65,
            "cup": 90,
            "block": 95,
            "average": 78
          },
          "detail": "Action features, with future gradients stopped",
          "ours": false
        },
        {
          "name": "D. Full FDM",
          "family": "vgm",
          "values": {
            "drawer": 70,
            "box": 75,
            "shuttlecock": 70,
            "cup": 100,
            "block": 100,
            "average": 83
          },
          "detail": "Action features + future-loss gradients",
          "ours": true
        }
      ]
    },
    "efficiency": {
      "label": "Efficiency",
      "title": "Future supervision at training, action-only at inference",
      "source": "Table 5",
      "href": "https://arxiv.org/pdf/2610.13117#page=10",
      "defaultMetric": "latency",
      "note": "Matched comparisons within each backbone family. Training uses 8 H100 GPUs; inference latency is measured on 1 H100, per action chunk.",
      "metrics": [
        {
          "key": "latency",
          "label": "Inference latency",
          "unit": "ms/chunk",
          "digits": 2,
          "max": 500,
          "higher": false
        },
        {
          "key": "step",
          "label": "Training step time",
          "unit": "s/step",
          "digits": 4,
          "max": 1,
          "higher": false
        },
        {
          "key": "throughput",
          "label": "Training throughput",
          "unit": "samples/s",
          "digits": 2,
          "max": 300,
          "higher": true
        },
        {
          "key": "memory",
          "label": "Peak training memory",
          "unit": "GiB/GPU",
          "digits": 2,
          "max": 80,
          "higher": false
        }
      ],
      "rows": [
        {
          "name": "π₀.₅",
          "family": "vlm",
          "values": {
            "latency": 203.77,
            "step": 0.4436,
            "throughput": 288.58,
            "memory": 67.08,
            "overhead": null,
            "speedup": 1,
            "reduction": 0
          },
          "detail": "Matched baseline",
          "ours": false
        },
        {
          "name": "FDM w/ VLM",
          "family": "vlm",
          "values": {
            "latency": 202.98,
            "step": 0.6305,
            "throughput": 203.01,
            "memory": 72.04,
            "overhead": 42.14,
            "speedup": 1,
            "reduction": 0.39
          },
          "detail": "Our model",
          "ours": true
        },
        {
          "name": "FastWAM-Joint",
          "family": "vgm",
          "values": {
            "latency": 446.55,
            "step": 0.9754,
            "throughput": 131.23,
            "memory": 66.8,
            "overhead": null,
            "speedup": 1,
            "reduction": 0
          },
          "detail": "Matched baseline",
          "ours": false
        },
        {
          "name": "FDM w/ VGM",
          "family": "vgm",
          "values": {
            "latency": 253.45,
            "step": 0.9729,
            "throughput": 131.56,
            "memory": 66.8,
            "overhead": -0.26,
            "speedup": 1.76,
            "reduction": 43.24
          },
          "detail": "Our model",
          "ours": true
        }
      ],
      "insights": {
        "latency": {
          "vlm": "Inference latency stays comparable to π₀.₅.",
          "vgm": "1.76× speedup; 43.24% lower latency than FastWAM-Joint."
        },
        "step": {
          "vlm": "The paper reports 42.14% more training time per step.",
          "vgm": "The paper reports 0.26% less training time per step."
        },
        "throughput": {
          "vlm": "Future supervision reduces training throughput in the VLM comparison.",
          "vgm": "Training throughput remains comparable to FastWAM-Joint."
        },
        "memory": {
          "vlm": "Future supervision adds 4.96 GiB of peak memory per GPU.",
          "vgm": "Both methods use 66.80 GiB of peak memory per GPU."
        }
      },
      "tableMetrics": [
        {
          "key": "latency",
          "label": "Inference latency",
          "unit": "ms/chunk",
          "digits": 2,
          "max": 500,
          "higher": false
        },
        {
          "key": "step",
          "label": "Training step time",
          "unit": "s/step",
          "digits": 4,
          "max": 1,
          "higher": false
        },
        {
          "key": "throughput",
          "label": "Training throughput",
          "unit": "samples/s",
          "digits": 2,
          "max": 300,
          "higher": true
        },
        {
          "key": "memory",
          "label": "Peak training memory",
          "unit": "GiB/GPU",
          "digits": 2,
          "max": 80,
          "higher": false
        },
        {
          "key": "overhead",
          "label": "Training overhead",
          "unit": "%",
          "digits": 2
        },
        {
          "key": "speedup",
          "label": "Inference speedup",
          "unit": "×",
          "digits": 2
        },
        {
          "key": "reduction",
          "label": "Latency reduction",
          "unit": "%",
          "digits": 2
        }
      ]
    }
  }
}