{
  "schema_version": "forcebench-paper-results-v1",
  "record_type": "submitted_paper_results",
  "is_sample": false,
  "source": {
    "paper_title": "ForceBench: Benchmarking Force Integration for Robotic Manipulation",
    "publication_status": "submitted",
    "submitted_date": "2026-09-15",
    "tables": {
      "simulation": "Table I",
      "real_world": "Table II"
    },
    "description": "Values transcribed from the submitted manuscript. Submission is not acceptance or publication."
  },
  "metric": {
    "name": "task_success_rate",
    "unit": "percent",
    "mean": "Unweighted arithmetic mean over the tasks in each study, displayed as reported to two decimal places.",
    "uncertainty": "No confidence intervals, significance tests, or per-training-seed estimates are reported in these tables."
  },
  "methods": [
    {
      "id": "pi05_no_force",
      "name": "π₀.₅ (No Force)",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "token_fusion",
      "name": "Token Fusion",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "prompt_fusion",
      "name": "Prompt Fusion",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "cross_attention",
      "name": "Cross-Attention",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "token_fusion_pred",
      "name": "Token Fusion + Pred",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "temporal_prompt",
      "name": "Temporal Prompt",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "dual_path_fusion",
      "name": "Dual-Path Fusion",
      "comparison_group": "controlled_pi05"
    },
    {
      "id": "pi0",
      "name": "π₀",
      "comparison_group": "additional_reference"
    },
    {
      "id": "forcevla",
      "name": "ForceVLA",
      "comparison_group": "additional_reference"
    },
    {
      "id": "forceflow",
      "name": "ForceFlow",
      "comparison_group": "additional_reference"
    }
  ],
  "simulation": {
    "source_table": "Table I",
    "tasks": [
      {
        "id": "weight_sort",
        "name": "Weight Sort",
        "category": "Physical-property inference"
      },
      {
        "id": "com_sort",
        "name": "CoM Sort",
        "category": "Physical-property inference"
      },
      {
        "id": "faucet_turn",
        "name": "Faucet Turn",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "faucet_direction",
        "name": "Faucet Direction",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "stove_turn",
        "name": "Stove Turn",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "stove_press",
        "name": "Stove Press",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "drawer_pull",
        "name": "Drawer Pull",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "nut_thread",
        "name": "Nut Thread",
        "category": "Contact-rich assembly"
      }
    ],
    "episodes_per_task_method": 200,
    "total_evaluation_episodes": 16000,
    "protocol": {
      "training_instance_seeds": {
        "start": 0,
        "end": 199,
        "inclusive": true
      },
      "evaluation_seeds": {
        "start": 300,
        "end": 499,
        "inclusive": true
      },
      "physics_rate_hz": 60,
      "control_rate_hz": 15,
      "max_control_steps": 1500,
      "max_simulated_seconds": 100,
      "separate_policy_per_task": true,
      "controlled_pi05_training_steps": 10000,
      "controlled_pi05_batch_size": 32,
      "action_chunks": [
        {
          "method_ids": [
            "pi05_no_force",
            "token_fusion",
            "prompt_fusion",
            "cross_attention",
            "token_fusion_pred",
            "temporal_prompt",
            "dual_path_fusion",
            "pi0",
            "forcevla"
          ],
          "predicted_steps": 50,
          "executed_steps": 50
        },
        {
          "method_ids": [
            "forceflow"
          ],
          "predicted_steps": 64,
          "executed_steps": 32
        }
      ],
      "notes": [
        "The seven pi0.5 configurations share the backbone and task demonstrations.",
        "The shared-backbone comparison includes changes in representation, normalization, capacity, and insertion location. Temporal designs serve as separate reference points; the predictive variant retains the Token Fusion input pathway.",
        "Additional reference systems are shown separately; their training and execution settings are not identical to the shared-backbone pi0.5 comparison.",
        "Faucet Direction uses the same lower-bound-only angle criterion for all ten systems.",
        "The control rate is not the policy-query rate."
      ]
    },
    "results": [
      {
        "method_id": "pi05_no_force",
        "success_rates_percent": [
          54.0,
          45.5,
          9.5,
          31.5,
          8.5,
          1.0,
          4.5,
          31.5
        ],
        "mean_success_percent": 23.25
      },
      {
        "method_id": "token_fusion",
        "success_rates_percent": [
          50.5,
          50.0,
          12.0,
          28.5,
          6.0,
          2.5,
          13.0,
          26.0
        ],
        "mean_success_percent": 23.56
      },
      {
        "method_id": "prompt_fusion",
        "success_rates_percent": [
          88.0,
          56.5,
          21.0,
          41.5,
          16.0,
          2.5,
          23.0,
          34.0
        ],
        "mean_success_percent": 35.31
      },
      {
        "method_id": "cross_attention",
        "success_rates_percent": [
          54.5,
          51.0,
          6.5,
          33.0,
          5.5,
          2.5,
          6.5,
          17.5
        ],
        "mean_success_percent": 22.13
      },
      {
        "method_id": "token_fusion_pred",
        "success_rates_percent": [
          78.0,
          48.5,
          13.5,
          32.5,
          6.0,
          1.0,
          4.0,
          26.5
        ],
        "mean_success_percent": 26.25
      },
      {
        "method_id": "temporal_prompt",
        "success_rates_percent": [
          90.0,
          69.5,
          17.5,
          31.0,
          7.0,
          21.5,
          2.5,
          22.0
        ],
        "mean_success_percent": 32.63
      },
      {
        "method_id": "dual_path_fusion",
        "success_rates_percent": [
          99.5,
          72.0,
          22.0,
          30.0,
          17.5,
          19.0,
          4.5,
          16.5
        ],
        "mean_success_percent": 35.13
      },
      {
        "method_id": "pi0",
        "success_rates_percent": [
          61.5,
          39.0,
          3.5,
          27.0,
          7.0,
          5.0,
          7.5,
          36.0
        ],
        "mean_success_percent": 23.31
      },
      {
        "method_id": "forcevla",
        "success_rates_percent": [
          35.5,
          45.5,
          12.5,
          14.0,
          10.0,
          12.0,
          20.0,
          26.0
        ],
        "mean_success_percent": 21.94
      },
      {
        "method_id": "forceflow",
        "success_rates_percent": [
          51.5,
          3.0,
          0.0,
          0.5,
          0.0,
          0.0,
          0.0,
          0.0
        ],
        "mean_success_percent": 6.88
      }
    ]
  },
  "real_world": {
    "source_table": "Table II",
    "tasks": [
      {
        "id": "weight_sort",
        "name": "Weight Sort",
        "category": "Physical-property inference"
      },
      {
        "id": "plate_retrieve",
        "name": "Plate Retrieve",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "stamp_press",
        "name": "Stamp Press",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "plug_insert",
        "name": "Plug Insert",
        "category": "Resistance-dependent interaction"
      },
      {
        "id": "block_insert",
        "name": "Block Insert",
        "category": "Contact-rich assembly"
      }
    ],
    "trials_per_task_method": 20,
    "total_evaluation_trials": 800,
    "protocol": {
      "demonstrations_per_task_method": 200,
      "training_steps": 10000,
      "batch_size": 64,
      "same_initial_conditions_across_methods": true,
      "robot": "Franka Research 3",
      "sensor": "Six-axis wrist force/torque sensor"
    },
    "results": [
      {
        "method_id": "pi05_no_force",
        "success_rates_percent": [
          30.0,
          20.0,
          40.0,
          60.0,
          50.0
        ],
        "mean_success_percent": 40.0
      },
      {
        "method_id": "token_fusion",
        "success_rates_percent": [
          45.0,
          40.0,
          50.0,
          60.0,
          10.0
        ],
        "mean_success_percent": 41.0
      },
      {
        "method_id": "prompt_fusion",
        "success_rates_percent": [
          70.0,
          40.0,
          65.0,
          80.0,
          60.0
        ],
        "mean_success_percent": 63.0
      },
      {
        "method_id": "cross_attention",
        "success_rates_percent": [
          60.0,
          30.0,
          40.0,
          50.0,
          55.0
        ],
        "mean_success_percent": 47.0
      },
      {
        "method_id": "token_fusion_pred",
        "success_rates_percent": [
          45.0,
          60.0,
          60.0,
          45.0,
          35.0
        ],
        "mean_success_percent": 49.0
      },
      {
        "method_id": "temporal_prompt",
        "success_rates_percent": [
          80.0,
          45.0,
          65.0,
          75.0,
          35.0
        ],
        "mean_success_percent": 60.0
      },
      {
        "method_id": "dual_path_fusion",
        "success_rates_percent": [
          95.0,
          45.0,
          70.0,
          85.0,
          20.0
        ],
        "mean_success_percent": 63.0
      },
      {
        "method_id": "forcevla",
        "success_rates_percent": [
          50.0,
          0.0,
          40.0,
          70.0,
          5.0
        ],
        "mean_success_percent": 33.0
      }
    ]
  }
}