{
  "article": "/research/so101-rl-to-lerobot-dataset/",
  "reported_as_of": "2026-09-18",
  "scope": "Simulation only; no real-robot VLA transfer result is claimed.",
  "evaluation_protocol": {
    "episodes_per_batch": 20,
    "base_seed": 12346,
    "reset_seed": "base_seed + episode_index",
    "starts": "Original-task randomized starts, matched across batches",
    "actions": "Deterministic learned policy only; no motion expert or action overrides",
    "physics_hz": 120,
    "collision_stack_bytes": 134217728,
    "clean_release": "Strict success plus no post-release contact >= 1 N and no jaw-target drawdown > 0.10 rad",
    "limitations": "One matched set of 20 starts, not multiple training seeds or a reliability guarantee. Clean-release success is stricter than task termination."
  },
  "checkpoints": {
    "baseline": {
      "variant": "stage_scaled_v1",
      "filename": "model_999.pt",
      "sha256": "88da5bfcd46afcf2512b22daead281ffb5fd57450346600c8987b81af47c8140"
    },
    "final": {
      "variant": "clean_release_v1",
      "filename": "model_999.pt",
      "sha256": "d88a22364832cc757132185081962d811364c6005c8ffeccdad20d89415b3e9c",
      "additional_fine_tuning_updates": 1000
    }
  },
  "evaluations": [
    {
      "policy": "baseline", "policy_hz": 60, "completed_episodes": 20,
      "secure_grasp": 20, "controlled_lift": 20, "strict_success": 9,
      "clean_release_success": 0, "timeout": 11, "drop": 0,
      "palm_strike": 0, "post_release_recontact": 3
    },
    {
      "policy": "final", "policy_hz": 60, "completed_episodes": 20,
      "secure_grasp": 20, "controlled_lift": 20, "strict_success": 20,
      "clean_release_success": 19, "timeout": 0, "drop": 0,
      "palm_strike": 0, "post_release_recontact": 0
    },
    {
      "policy": "baseline", "policy_hz": 30, "completed_episodes": 20,
      "secure_grasp": 9, "controlled_lift": 9, "strict_success": 2,
      "clean_release_success": 0, "timeout": 18, "drop": 0,
      "palm_strike": 2, "post_release_recontact": 2
    },
    {
      "policy": "final", "policy_hz": 30, "completed_episodes": 20,
      "secure_grasp": 13, "controlled_lift": 13, "strict_success": 12,
      "clean_release_success": 9, "timeout": 8, "drop": 0,
      "palm_strike": 3, "post_release_recontact": 0
    }
  ],
  "dataset": {
    "format": "LeRobotDataset v3",
    "accepted_episodes": 1000,
    "completed_attempts": 3008,
    "frames": 86832,
    "fps": 30,
    "policy_hz": 30,
    "camera_resolution": [640, 480],
    "camera_streams": ["top", "wrist"],
    "state_dimensions": 6,
    "action_dimensions": 6,
    "all_accepted_pass_strict_success": true,
    "all_accepted_pass_clean_release": true,
    "limitations": "Success-filtered collection, not an unbiased policy evaluation. Incomplete crash-interrupted attempts are excluded from the completed-attempt count. Real robot action calibration and sim-to-real transfer remain unverified."
  }
}