myconnects/motif_actions

MotIF-Actions

Robot action-space companion to the MotIF-1K dataset. The public MotIF release exposes only the 2D end-effector pixel trajectory; MotIF-Actions adds the on-board Stretch proprioception for the subset of demonstrations that were logged with joint states.

Built directly from the raw teleoperation recordings, in RLDS / TFDS format.

Coverage

	trajectories
Joint state (full)	237
+ task / motion language labels	202
+ 2D EEF on the camera frame	175
+ 3D EEF (precomputed)	35

(The 3D EEF is provided where it was logged at capture time; the other trajectories can be forward-kinematics'd from the joint state using the Stretch URDF.)

Per-step fields

field	shape	description
observation.image	(240, 320, 3) uint8	RGB sideview frame — the raw uncropped camera frame, downscaled 0.5× from the native 480×640
observation.state	(10,) float32	joint positions: [base_x, base_y, lift, arm, wrist_yaw, wrist_pitch, wrist_roll, gripper, head_pan, head_tilt]
observation.joint_velocity	(10,) float32	joint velocities (same order)
observation.eef_2d	(2,) float32	end-effector [x, y] pixel on the camera frame (zeros if eef_2d_valid is False)
observation.eef_2d_valid	bool	whether eef_2d is real
observation.eef_3d	(3,) float32	end-effector [x, y, z] in the robot 3D frame, meters (zeros if eef_3d_valid is False)
observation.eef_3d_valid	bool	whether eef_3d is real
action	(8,) float32	delta of controllable joints to the next step: [base_x, base_y, lift, arm, wrist_yaw, wrist_pitch, wrist_roll, gripper]
language_instruction	text	task instruction (e.g. "shake boba")
language_instruction_2	text	motion description (e.g. "move right and left 3 times")

Standard RLDS step flags (is_first, is_last, is_terminal, reward) are included; reward is 1.0 on the last step.

Episode metadata: file_path, traj_idx, session, has_eef_2d, has_eef_3d.

Coordinate frames

• eef_2d is expressed in native camera-frame pixels (480×640) — not the cropped frame used by the public MotIF visualizations. The stored observation.image is downscaled 0.5×, so to overlay eef_2d on it, multiply by episode_metadata.image_scale (= 0.5). episode_metadata.orig_image_hw gives the native [H, W].
• eef_3d and joint positions are in the robot's frame (Stretch conventions: base_x/base_y planar translation in meters, lift/arm prismatic in meters, wrists in radians, head_pan/head_tilt in radians).

Loading

import tensorflow_datasets as tfds
b = tfds.builder_from_directory("data/1.0.0")   # from a local clone
ds = b.as_dataset(split="train")
for ep in ds:
    for step in ep["steps"]:
        state  = step["observation"]["state"]      # (10,)
        action = step["action"]                     # (8,)
        eef3d  = step["observation"]["eef_3d"]      # (3,)

Citation

@inproceedings{motif2024,
  title={MotIF: Motion Instruction Fine-tuning},
  year={2024}
}