{
  "kind": "estimated_local_displacement_projection",
  "episode": 14,
  "camera": 2,
  "image_size": [
    1280,
    720
  ],
  "matrix_xyz1_to_uv": [
    [
      -608.720259038727,
      -568.8671113862729
    ],
    [
      -1195.6168388923634,
      389.0865933515354
    ],
    [
      -155.75148841128336,
      -1219.1341832519315
    ],
    [
      -368.0972560698997,
      538.372073029908
    ]
  ],
  "video_to_state_offset_seconds": -0.2799999999999999,
  "anchor_image_px": [
    535.0,
    190.0
  ],
  "anchor_description": "Manually selected midpoint between visible jaws in first frame, not a calibrated TCP location.",
  "tracking": "Pyramidal Lucas-Kanade flow on rigid pneumatic-body corners; median displacement of persistent features; forward-backward error <0.5 px.",
  "initial_features": 60,
  "final_features": 60,
  "split": "Fit on even-numbered two-second video blocks; validate on odd blocks. Offset selected on training blocks only.",
  "held_out_rmse_px": 1.763675712377981,
  "held_out_p95_px": 4.114010588009238,
  "held_out_max_px": 5.277670794827665,
  "training_rmse_px": 1.5902225597575004,
  "fitted_xyz_range_m": [
    [
      -0.010744,
      -0.753436,
      0.016409
    ],
    [
      0.039216,
      -0.643093,
      0.049985
    ]
  ],
  "limitations": "Local affine displacement estimate for this fixed view and nearly fixed gripper orientation. The origin is illustrative. No measured intrinsics/extrinsics are available. Intermediate noisy positions extend beyond the fitted motion volume and are extrapolated. Not a camera calibration, executed-plan reconstruction, or metric validation of predicted targets."
}