action_mapping.json

{
  "action_space": "droid_raw_joint_position_8d",
  "droid_layout": [
    "joint_0",
    "joint_1",
    "joint_2",
    "joint_3",
    "joint_4",
    "joint_5",
    "joint_6",
    "gripper"
  ],
  "action_slots_in_xwam_14d": [
    0,
    1,
    2,
    3,
    4,
    5,
    7,
    6
  ],
  "proprio_slots_in_xwam_16d": [
    0,
    1,
    2,
    3,
    4,
    5,
    6,
    7
  ],
  "normalization": "y = clip(2*(x-q01)/(q99-q01) - 1, -1, 1); the gripper channel is then NEGATED",
  "gripper_convention_after_normalization": "+1 = open, -1 = closed (X-WAM convention)",
  "gripper_convention_raw_droid": "0 = open, 1 = closed",
  "q01": [
    -0.9290000200271606,
    -0.878000020980835,
    -0.9549999833106995,
    -2.7760000228881836,
    -1.8550000190734863,
    1.156999945640564,
    -2.0980000495910645,
    0.0
  ],
  "q99": [
    1.0,
    1.4529999494552612,
    0.8420000076293945,
    -0.43299999833106995,
    1.7330000400543213,
    3.5299999713897705,
    2.242000102996826,
    1.0
  ],
  "video_size_hw": [
    192,
    320
  ],
  "views": [
    "observation.images.exterior_1_left",
    "observation.images.exterior_2_left",
    "observation.images.wrist_left"
  ],
  "fps_video": 3.75,
  "fps_action": 15.0,
  "depth_root": "/iopsstor/scratch/cscs/hongsu/imbw/droid_depth_da3",
  "depth_normalization": "DA3 inverse depth (near=bright), per (episode, camera) robust affine to [0,1] at generation time; then per-view min/max over the window to [-1,1]"
}