Datasets:

Tsaochengyu
/

0916_kitchen_datasegment

This dataset contains the drawer-opening portion extracted from all 104 official RoboCasa DeliverStraw pretraining human demonstrations. The source is pretrain/composite/DeliverStraw/20250723/lerobot.tar from NVIDIA's PhysicalAI-Robotics-Manipulation-Kitchen-Demos release.

The complete source task is: take a straw from the drawer in front and place it inside the glass cup on the dining counter. This derived dataset contains only the drawer-opening stage, not the subsequent straw pickup and placement stages.

Segmentation

For each source episode, the exact recorded MuJoCo state and episode model are restored frame by frame. The segment begins five frames before the first physical contact between a gripper finger collision geometry and the target drawer handle. It ends five frames after the drawer first reaches 95% opening.

The event boundaries are determined from simulation contact and drawer state, not from image estimates or action replay. audit.jsonl records the source episode, source frame range, contact frame, opening frame, and drawer name.

Handle Frame

The moving Handle Frame is evaluated independently at every frame:

origin: center of the moving drawer handle;
x axis: drawer width direction;
y axis: drawer depth direction;
z axis: drawer up direction.

Its orientation follows the moving drawer body. For a world-frame pose (p_W, R_W_E) and Handle Frame (o_W, R_W_H), the conversion is:

p_H   = R_W_H^T (p_W - o_W)
R_H_E = R_W_H^T R_W_E

Default Model Interface

The default 16D observation.state is:

[base_position_in_handle_frame(3),
 base_rotation_quaternion_xyzw_in_handle_frame(4),
 eef_position_in_handle_frame(3),
 eef_rotation_quaternion_xyzw_in_handle_frame(4),
 gripper_qpos(2)]

The default 15D action is:

[base_motion(4), control_mode(1),
 commanded_eef_position_in_handle_frame(3),
 commanded_eef_x_axis_in_handle_frame(3),
 commanded_eef_y_axis_in_handle_frame(3),
 gripper(1)]

The commanded end-effector target is reconstructed by restoring the current simulation state and applying the original action to the RoboCasa controller. It is then expressed in the current Handle Frame. All extracted frames use arm mode: the four base-action values are zero and control_mode is -1.

The 6D rotation fields store the first two columns of the rotation matrix and are recommended for behavior cloning because they avoid rotation-vector branch discontinuities.

Preserved and Additional Fields

observation.state.original: original official 16D proprioceptive state.
observation.handle_relative_state: explicit copy of the default 16D state.
observation.handle_relative_state_6d: 20D base and EEF state with 6D rotations.
observation.handle_relative_eef_state: compact 8D EEF state with rotvec.
observation.handle_relative_eef_state_6d: compact 11D EEF state.
action.original: original official 12D normalized controller action.
action.robot_base_delta: explicit copy of the original 12D action.
action.handle_relative_target: 12D Handle Frame target with rotvec.
action.handle_relative_target_6d: explicit copy of the default 15D action.
action.handle_relative_eef_target: compact 7D EEF target with rotvec.
action.handle_relative_eef_target_6d: compact 10D EEF target.
observation.handle_relative_achieved_pose: next-frame achieved EEF pose.
observation.handle_relative_achieved_pose_6d: its 6D rotation variant.
source_episode_index and source_frame_index: source traceability.

The achieved-pose fields are validation labels rather than current-frame inputs. The three camera streams remain in their native camera views and are only temporally clipped. Each states.npz contains the unmodified MuJoCo simulation states for the same clipped frame range.

Dataset Contents

104 valid episodes and 0 invalid episodes
7,196 frames at 20 FPS
104 per-episode Parquet files
312 clipped MP4 files from three cameras
104 clipped states.npz files
104 original per-episode model.xml.gz files
segment length: 34 to 348 frames, mean 69.19 frames

Validation

Every episode was checked for aligned Parquet, simulation-state, and video frame counts; continuous global and per-episode indices; exactly one terminal flag; finite values and expected dimensions for all state/action representations; and orthonormal 6D rotation axes. The maximum observed 6D orthogonality error was 6.66e-16.

Source and License

Source dataset: https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Manipulation-Kitchen-Demos

Source file: pretrain/composite/DeliverStraw/20250723/lerobot.tar

The source dataset is released under CC BY 4.0. This derived dataset retains that license and source attribution.

Downloads last month: 203