{
  "schema_version": "anchor.robodojo.long-horizon-comparison.v1",
  "generated_at": "2026-07-21",
  "claim": "Two official fill_egg_holder teleoperation demonstrations have similar closed-holder final appearances, while step-level evidence reveals a 2.73x duration gap and materially different execution cost.",
  "source": {
    "dataset": "RoboDojo-Benchmark/RoboDojo",
    "dataset_revision": "1a3c4c334aef294c31d7a0190d8d6dff68df78e0",
    "dataset_license_metadata": "apache-2.0",
    "task_id": "fill_egg_holder",
    "robot": "arx_x5",
    "category": "Long-Horizon",
    "data_source": "Teleop",
    "instruction": "Place the four eggs from the basket into the egg holder, then close the lid.",
    "task_definition": "https://robodojo-benchmark.com/doc/sim-tasks/fill-egg-holder/",
    "dataset_path": "https://huggingface.co/datasets/RoboDojo-Benchmark/RoboDojo/tree/main/data/RoboDojo/fill_egg_holder/arx_x5",
    "scoring_code": "https://github.com/RoboDojo-Benchmark/RoboDojo/blob/main/task/RoboDojo/tasks/fill_egg_holder.py"
  },
  "comparison_design": {
    "type": "cohort_examples",
    "official_demonstrations": true,
    "controlled_same_seed_policy_ab": false,
    "official_score_rerun": false,
    "robustness_claim": "none; two-example investor demonstration"
  },
  "episodes": {
    "efficient": {
      "local_episode": 2,
      "lerobot_global_episode": 702,
      "frames": 270,
      "fps": 25,
      "video_duration_seconds": 10.8,
      "trajectory_duration_seconds": 10.76,
      "action_path_l2": 24.000882,
      "action_delta_sign_change_reversals": 850,
      "low_activity_steps_l2_lt_0_01": 3,
      "low_activity_step_coverage_denominator": 269,
      "peak_action_delta_l2": 0.345166,
      "camera_frame_sync": {
        "head": 270,
        "left_wrist": 270,
        "right_wrist": 270
      },
      "final_appearance": "closed-holder appearance in final head frame; official score unscored"
    },
    "detour": {
      "local_episode": 54,
      "lerobot_global_episode": 754,
      "frames": 736,
      "fps": 25,
      "video_duration_seconds": 29.44,
      "trajectory_duration_seconds": 29.4,
      "action_path_l2": 29.78486,
      "action_delta_sign_change_reversals": 2533,
      "low_activity_steps_l2_lt_0_01": 123,
      "low_activity_step_coverage_denominator": 735,
      "peak_action_delta_l2": 0.197232,
      "camera_frame_sync": {
        "head": 736,
        "left_wrist": 736,
        "right_wrist": 736
      },
      "final_appearance": "closed-holder appearance in final head frame; official score unscored"
    }
  },
  "relative_result": {
    "duration_ratio_detour_vs_efficient": 2.73,
    "efficient_video_time_reduction_percent_vs_detour": 63.32,
    "efficient_action_path_reduction_percent_vs_detour": 19.42,
    "efficient_reversal_reduction_percent_vs_detour": 66.44,
    "efficient_low_activity_step_reduction_percent_vs_detour": 97.56,
    "efficient_low_activity_rate_reduction_percent_vs_detour": 93.34
  },
  "metric_notes": {
    "video_duration": "All three camera streams are synchronized by frame count within each episode at 25 FPS.",
    "action_path_l2": "Sum of L2 distances between adjacent 14-dimensional action vectors; it is not Cartesian distance in meters.",
    "reversals": "Count of sign changes in per-joint action deltas; a motion-correction proxy, not a semantic retry count.",
    "low_activity_steps": "Adjacent 14-dimensional action delta L2 below 0.01; an activity proxy, not proof that the robot physically stopped.",
    "official_score": "unscored; the RoboDojo evaluator was not rerun for these demonstrations"
  },
  "excluded_claims": [
    "The two episodes are not a controlled same-seed, same-layout, two-policy A/B.",
    "The page does not claim that either episode received an official RoboDojo score of 100.",
    "Direction reversals and low-activity steps are proxies, not human-labeled retries or stalls.",
    "Two examples do not establish cross-seed robustness."
  ]
}