exaFLOPs09/sink10k_mug_g2_grasp

sink10k-mug: phase 2, grasp

The second phase only -- closing on the mug and lifting it, cut out of the complete demonstrations.

An RB-Y1 mobile manipulator picking a mug off a kitchen worktop and putting it in the kitchen sink, generated in simulation (NVIDIA Isaac Lab). Nothing in this dataset was recorded on hardware.

What is in it

demonstrations	9,801
frames	2,621,199
kitchens	122
kitchen scenes (kitchen x layout variant)	223
control and video rate	20 Hz
cameras	3
format	LeRobot v3.0

Prompt, for every frame of this dataset: Grasp the mug and lift it.

Cameras

All three are 320x240 RGB, h264, at 20 Hz -- the same rate as the control stream.

key	where it is
observation.images.front	head camera, mounted on the robot's head link and pitched 50 degrees down
observation.images.wrist_left	left wrist
observation.images.wrist_right	right wrist

Robot signals

Read this before you train on a column. action.joint is a frozen snapshot: it is written once per episode and never changes again, so every frame of a demonstration carries the same 29 numbers. It is useless as a learning target and you must not train on it. joint_angles is the joint signal in this dataset. action.joint is kept only so the schema still matches the exports it was merged from.

key	shape	what it is
observation.state	23	[position(3), 6D rotation(6), gripper(1)] per arm, then the base twist (v_x, v_y, omega)
action	23	the commanded version of the same layout
joint_angles	24	measured joint angles -- the joint signal
action.joint	29	commanded joint targets. Constant within an episode. Do not train on it.

A second warning about observation.state, because it would bite silently: the exports this was merged from carried the wrong channel names for it, and the names in this repository have been corrected. The exports labelled the two grippers as indices 18 and 19, which shifted every label from index 9 to 18 one place off the channel it names -- r_x named the gripper. The real layout is per-arm, with each arm's gripper last in its own block, and the two gripper channels are the way round the corrected names now say: index 9 is the RIGHT gripper and index 19 is the left. This was established by measurement, not assumption -- index 10:13 tracks the correctly named action[10:13] at r > 0.98, index 9 tracks the gripper_finger_r1 joint angle, and index 19 lands where the same units put a resting left gripper. Only the labels were changed; not one recorded number was touched. If you have earlier copies of these exports, their observation.state names are wrong.

Each frame also carries kitchen_num, kitchen_sub_num, kitchen_type, initial_pose, is_first, is_last and subtask_index. The exact dtypes and the per-joint names are in meta/info.json.

One more name that does not say what it holds: initial_pose is not the initial pose. It carries the robot base's (x, y, qw, qx, qy, qz) AT EVERY FRAME and changes on every one of them -- a 596-frame episode has 596 distinct values. Read it as the base's live pose, not as a per-episode constant. The name is left as it is here because it is the parquet column's own name; renaming the label without rewriting every row would only move the problem.

The four phases

Every demonstration is also published cut into four phases. The cut is by frame, so the four phases partition each episode exactly and their frame counts sum to the complete dataset's. All five datasets hold the same 9,801 demonstrations; only the frame ranges and the prompt differ.

phase	repository	prompt	episodes	frames
approach	exaFLOPs09/sink10k_mug_g1_approach	Move to the mug.	9,801	3,567,659
grasp	exaFLOPs09/sink10k_mug_g2_grasp	Grasp the mug and lift it.	9,801	2,621,199
carry	exaFLOPs09/sink10k_mug_g3_carry	Carry the mug to the sink.	9,801	4,426,741
place	exaFLOPs09/sink10k_mug_g4_place	Put the mug in the sink and return the arm home.	9,801	1,015,587

The complete task is exaFLOPs09/sink10k_mug, with the prompt Grasp the mug and put it in the sink..

All five datasets

• exaFLOPs09/sink10k_mug
• exaFLOPs09/sink10k_mug_g1_approach
• exaFLOPs09/sink10k_mug_g2_grasp -- this dataset
• exaFLOPs09/sink10k_mug_g3_carry
• exaFLOPs09/sink10k_mug_g4_place

Reach directness, and what was not filtered

How straight the right hand's path to the mug is varies from demonstration to demonstration. Measured over all 9,801 episodes -- not a sample -- over the arm.grasp step alone, as the length of the right end effector's path divided by the straight line from where that path starts to where it ends. A ratio of 1.0 is a perfectly direct reach.

reach_detour_ratio
median	1.1913
75th percentile	1.4085
90th percentile	1.7478
99th percentile	2.3558
maximum	12.7992
episodes over 2.0	404
episodes over 10.0	1

The reach itself is 0.4209 m long at the median. No detour RATIO is published for the wider grasp phase, on purpose: that phase is reach, close and lift, and the lift returns the hand towards where it started, so the straight line in the denominator collapses and a ratio there would measure how closed the loop is rather than how direct the path was. For that window the file gives the honest quantities instead -- grasp_path_len_m, the distance actually travelled (0.9559 m at the median), and grasp_straight_line_m, the net displacement.

No demonstration was filtered or dropped on this measure, or on any other measure of motion quality. The less direct reaches sit in the dataset alongside the direct ones, deliberately, so that the choice is yours and not one already made for you behind a threshold you cannot see.

The per-episode numbers ship with the data as episode_motion_stats.csv, one row per demonstration for all 9,801 of them: episode_index, kitchen, sub, then reach_frames, reach_path_len_m, reach_straight_line_m, reach_detour_ratio and reach_peak_right_arm_joint_speed_rad_s for the reach, and grasp_frames, grasp_path_len_m, grasp_straight_line_m and peak_right_arm_joint_speed_rad_s for the whole grasp phase (peak right-arm joint speed over the reach: median 0.9363 rad/s, 90th percentile 5.3364 rad/s, maximum 10.8871 rad/s). episode_index identifies the same demonstration in all five of these datasets, so a selection made against one applies to any of them.

A note on recovery

468 of the 9,801 demonstrations (24 of the 294 generation batches) were written to disk without their parquet footer and were rebuilt by re-reading the row groups. The rebuilt episodes carry the same columns and the same frame counts as the rest and passed the same checks; they are called out here because a rebuilt file is a fact about the data worth knowing, not because anything is known to be wrong with them.

How this was checked

Measured on the merged dataset, not asserted:

• 9,801 episodes and 2,621,199 frames, the same number in meta/info.json, in the parquet footers and in the parquet rows.
• episode_index runs 0..9,800 with no duplicates; index runs 0..2,621,198 in order with no duplicates.
• Per-episode lengths agree between meta/episodes and the data; frame_index restarts at 0 in every episode.
• One task string, Grasp the mug and lift it., referenced by every frame.
• 60 episodes compared value-for-value against the export they came from, half of them on a merge boundary.
• Every one of the 29,403 video windows checked against the duration of the file it points at; 882 merged video files checked against their source, 30 of them by sha256; and 180 real frames decoded from 60 episodes and compared pixel for pixel with the source.