Datasets:

unitreerobotics
/

R1_Dex1_Put_Pens_Into_Drawer

4 cameras, MP4 (mp4v, yuv420p), 480×640×3, 30 fps, no audio. The data is stored as videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4.

Key	Description
`observation.images.cam_left_high`	Left head/high camera
`observation.images.cam_right_high`	Right head/high camera
`observation.images.cam_left_wrist`	Left wrist camera
`observation.images.cam_right_wrist`	Right wrist camera

Observations (`observation.state`, flat 16-D)

A single flat float32[16] vector, no nested keys. Order:

Idx	Name	Description
0	`kLeftShoulderPitch`	Left arm
1	`kLeftShoulderRoll`
2	`kLeftShoulderYaw`
3	`kLeftElbow`
4	`kLeftWristRoll`
5	`kLeftWristPitch`
6	`kLeftWristYaw`
7	`kRightShoulderPitch`	Right arm
8	`kRightShoulderRoll`
9	`kRightShoulderYaw`
10	`kRightElbow`
11	`kRightWristRoll`
12	`kRightWristPitch`
13	`kRightWristYaw`
14	`kLeftDex1`	Left Dex1 gripper
15	`kRightDex1`	Right Dex1 gripper

Arm joints (idx 0–13) are joint positions in radians. Gripper channels (idx 14–15) are the raw Dex1 values (see Hand Convention).

Actions (`action`, flat 16-D)

Same layout and names as observation.state: float32[16], arm joints (idx 0–13) plus the two Dex1 gripper commands (idx 14–15). In this dataset the gripper command range is exactly [0.0, 5.0].

Bookkeeping

timestamp, frame_index, episode_index, index, task_index — each shape [1] (timestamp float32, the rest int64).

Hand Convention

This dataset uses Dex1 grippers (kLeftDex1 / kRightDex1, 1 channel each, at index 14/15 of both observation.state and action).

Dex1 Gripper (range 0.0 – 5.0)

Raw values are kept as-is; no normalization is applied at storage time.
Observed in this dataset: observation.state gripper ≈ [-0.04, 5.20], action gripper [0.0, 5.0].

Note: the conversion pipeline unifies open/close as 1 = open, 0 = closed for the 6-DOF hands (brainco/inspire); those hands are stored unnormalized too. The Dex1 single-DOF value is stored raw, so the two gripper extremes are 0 and 5.