JingwuLuo/ur3_sim_pick_and_place_cube

UR3 simulated pick and place cube (raw HDF5)

Simulated pick-and-place episodes on a replica of the dual-arm UR3 workcell, recorded in the HDF5 layout of the real datasets JingwuLuo/pick_and_place_cube_quick and JingwuLuo/pick_and_place_quick_new, plus a sim/ group with the simulator's ground truth. The workcell follows the real rig in its setting of 2026-11-03 (revision 88bfa62 of pick_and_place_quick_new): the top camera records 640 × 360 (16:9) at the pose fitted to that recording.

• 300 episodes, episode_0000.hdf5 … episode_0299.hdf5, one HDF5 file per episode.
• 44,043 steps at 30 Hz (24.5 min); episodes 2.6–10.1 s, median 4.5 s; 2.0 GiB.
• In every episode the right arm picks one of three cubes and places it on the white box, the plate or another cube; the left arm stays at its start pose. The table is never the target.
• The gripper is upright unless a pose needs a tilt: where the cube or the place is too far for an upright gripper, it leans by the smallest angle that reaches it (30 grasps and 22 releases, at most 8°), so the tilt is a function of the pose. Start and end poses are upright.
• The gripper's yaw turns like a teleoperator's: towards the cube's faces but not always square with them (up to 20° short), and back near the home yaw for the release, whatever the target's orientation.
• 72 episodes are perturbed: the arm is pushed 1–5 cm off its path during the approach or the transport, stops, and corrects from there. The pushed steps are tagged (command/right/source = "perturbation").
• 47 episodes have a deliberate late close and regrasp.
• Pick-to-place distances are spread evenly from 0.08 to 0.53 m (33–34 episodes per 5 cm bin).
• The top camera records 640 × 360 (16:9, horizontal field of view 69.4°); its view is turned by up to ±3° in pan and tilt per episode.
• The motions are scripted from the simulator's privileged state (object poses), not teleoperated and not produced by a learned policy.
• Only successful episodes are included (success check under Generation).
• No language annotation: language_instruction is the empty string in every episode, and the task column of episodes.csv is empty. The sim/task group names the arm, cube, source and target of each episode.

Sample videos (top camera on the left; left and right wrist cameras on the right):

video	episode
samples/episode_0019.mp4	pink cube from the table onto the plate, 0.21 m; grasped near the camera with the gripper tilted 8°
samples/episode_0247.mp4	pink cube from the plate onto the box, 0.34 m; picked at y −0.54 m, the outer edge of the right arm's region; released near the camera, tilted 7°
samples/episode_0000.mp4	orange cube from the table onto the box, 0.50 m; picked at y −0.53 m
samples/episode_0129.mp4	orange cube from the pink cube onto the plate, 0.39 m; pushed while carrying the cube; placed at y −0.54 m
samples/episode_0008.mp4	orange cube from the pink cube onto the box, 0.40 m; pushed during the approach, then a late close: the fingers close on air after the arm has lifted, then regrasp
samples/episode_0197.mp4	orange cube from the plate onto the blue cube, 0.45 m; grasp tilted 7°; late close and regrasp
samples/episode_0299.mp4	pink cube from the box onto the orange cube, 0.13 m; pushed during the approach
samples/episode_0036.mp4	blue cube from the table onto the pink cube, 0.11 m

The caption's second line gives the gripper tilt from vertical at the grasp and the release.

Quick start

import cv2, h5py

with h5py.File("episode_0000.hdf5", "r") as f:
    task = dict(f["sim/task"].attrs)            # arm, cube, source, target_kind, target, distance
    q = f["observations/right/joint_pos"][:]     # (T, 6) rad
    grip = f["command/right/gripper"][:]         # (T,) 0 open / 1 closed
    source = f["command/right/source"][:]        # (T,) b"script" | b"perturbation" | b"idle"
    pan, tilt = f["sim/top_camera_offset"][:]    # top camera turn of this episode, degrees
    i = 100
    top = cv2.imdecode(f["observations/images/top"][i, :f["observations/images/top_len"][i]], cv2.IMREAD_COLOR)

Contents

arm	right in all 300; no relays between the arms
pick-to-place distance (xy)	33–34 episodes per 5 cm bin from 0.08 to 0.53 m; median 0.31 m, p10 0.13 m, p90 0.48 m, max 0.53 m
perturbed	72: during the approach 23, while carrying 49
late close and regrasp	47 episodes (late_close in episodes.csv)
cubes	pink 91, orange 105, blue 104
gripper tilt	upright at 270 grasps and 278 releases; tilted grasps 1–8° (30), releases 1–8° (22)
overhead light shape	60 each of spot, pair, tube, panel, ring
top camera turn	pan and tilt each uniform in ±3°

Source → target (the cube's support before → after):

source \ target	cube	plate	box
table	34	34	34
cube	—	33	33
plate	33	—	33
box	33	33	—

Motion per episode:

quantity	p10	median	p90	max
duration (s)	3.4	4.5	7.0	10.1
peak flange speed (m/s)	0.38	0.48	0.60	0.66
peak joint speed (rad/s)	1.53	1.87	2.38	2.51

Turn of the fingers' closing direction between key poses (p50 / p75 / p95, degrees), against the right arm of the real recordings:

	this dataset	pick_and_place_quick_new 88bfa62	pick_and_place_quick_new 83f5758	pick_and_place_cube_quick
start → grasp	14 / 23 / 32	9 / 21 / 48	11 / 20 / 35	7 / 17 / 38
grasp → release	16 / 28 / 46	16 / 25 / 36	12 / 21 / 27	11 / 17 / 31
release → end	9 / 15 / 27	13 / 21 / 35	10 / 16 / 26	8 / 17 / 32

For comparison, the 80 teleoperated episodes of pick_and_place_quick_new revision 88bfa62 have a median duration of 9.1 s; over the steps at which the right flange moves faster than 0.01 m/s, its speed has a median of 0.13 m/s, p99 0.46 m/s and max 0.64 m/s.

Coordinate frame

Positions in sim/ are in the workcell frame:

• x points from the arms towards the top camera; y points towards the left arm; z is up from the floor.
• The table top is at z = 0.75 m. The arm bases are at (−0.30, ±0.36, 0.781) m (left at +y).
• The table spans x −0.275 to 0.685 m and y −0.95 to 0.65 m. A green backdrop cloth hangs behind the arms at x = −0.62 m.

observations/{left,right}/eef_pose is the flange (tool0) in that arm's own controller base frame, as in the real datasets.

Simulation

• Simulator: MuJoCo 3.17.0; physics at 600 Hz, recorded at 30 Hz.
• Robots: two UR3 arms (kinematics, inertias and meshes from UniversalRobots/Universal_Robots_ROS2_Description) with DH AG-105-145 grippers.
• Workcell geometry fitted to the real datasets:
  • arm base height 0.031 m above the table, from fingertip contacts in pick_and_place_cube_quick;
  • table edges behind the bases and on the left arm's side, from the top images of pick_and_place_quick_new revision 88bfa62;
  • top camera pose from cube and wrist-bracket correspondences in the 640 × 360 top images of pick_and_place_quick_new revision 88bfa62 (reprojection error median 5.3 px for the cubes): lens at (0.706, −0.054, 1.330) m, looking at (−0.019, −0.046, 0.75) m, vertical field of view 42.6° (focal length 462 px, the D435i colour camera's nominal value at this mode).
• Objects, with sizes and colours estimated from the real top images:
  • cubes: pink, orange, blue, 48 mm edge;
  • white box: 96 × 93 × 37 mm;
  • plate: 92 mm diameter, 10 mm rim, sloped rim around a 58 mm flat centre.
• Backdrop colour: its pixels in the top camera match those of pick_and_place_quick_new revision 83f5758 (median RGB about (16, 37, 25) in the simulation against (15, 35, 23), both with the top camera of 2026-11-02).
• Arm control: joint position servos with velocity feedforward, gravity-compensated as the real UR3 controller is.
• Gripper: grasps hold by contact friction only (no attachment). The opening target moves at 0.45 of the stroke per second while closing and 1.25 while opening, as measured on the real gripper (about 0.41 closing; 1.24 and 1.28 opening in pick_and_place_cube_quick and in revision 83f5758 of pick_and_place_quick_new). Open command 0.433, as on the real rig.
• Cameras:
  • top: 640 × 360 (16:9; vertical field of view 42.6°, horizontal 69.4°), the fitted pose above, turned per episode by a pan about the vertical through the lens and a tilt about the camera's horizontal axis, each uniform in ±3° (stored in sim/top_camera_offset); the objects are placed inside each episode's view.
  • left_wrist / right_wrist: 424 × 240, D405 vertical field of view 58°.
    • Mount: the D405 bracket of the gripper model. Both D405s sit on the table side of their grippers at the teleoperation home, as on the real rig. The left camera is the model's mount turned by 180° about the tool axis.
    • Orientation noise per episode: normal with σ 2°, clipped at 5° (stored in sim/wrist_camera_noise).
  • Images are JPEG, quality 90, from MuJoCo's OpenGL renderer with shadows. There is no depth.
• Lighting, one overhead set per episode in addition to a fixed headlight:
  • shape: spot (cone 30–60°), pair (0.4–1.2 m apart), tube (5 lights on a 0.4–1.2 m line), panel (2 × 3 grid, 0.3–0.9 m) or ring (6 lights, radius 0.15–0.5 m);
  • centre x −0.3 to 0.5 m, y ±0.5 m, z 1.7 to 2.5 m; tilt up to 25°;
  • colour: a shift of red against blue from 14 % cool to 6 % warm, plus 1.5 % per-channel jitter (the table's blue/red ratio in the top images spans 0.83–1.06; 0.885–1.052 in pick_and_place_quick_new revision 83f5758);
  • mean table irradiance 0.85–1.15 times the reference rig under which object and table colours were matched. A narrow spot cannot always reach that within the per-light limit: 16 episodes (11 of them spot) have a mean table irradiance below 0.8 times the reference, down to 0.22.

Generation

1. Cell. Every episode id gets a cell before collection: the cube's support kind, the target kind, the light shape, the distance bin and whether it is perturbed. Cells are assigned so that every factor's counts are even: 33–34 episodes per (support kind, target kind) pair and per distance bin, 60 per light shape; no (pair, shape, bin) combination repeats; a quarter of each bin is perturbed. An attempt counts only for its id's cell.
2. Arrangement.
   • The top camera's turn of the episode is drawn first.
   • The box and the plate stand on the table.
   • Each cube rests on the table (55 %), on the box, on the plate or on another cube (15 % each).
   • Positions cover both arms' reach inside the episode's top camera view (x −0.22 to 0.30 m, y ±0.56 m, at least 0.20 m from the arm bases), so objects also lie where only the left arm reaches.
3. Task.
   • The cube: a free cube on a support of the cell's kind that the right arm reaches.
   • The target: of the cell's kind, at a pick-to-place distance in the cell's bin, and reached by the right arm.
   • Reached: the gripper can be placed at the point's cube centre and 0.10 m above it, for one of the four yaws across the cube's faces, upright or leaning 15° towards or away from the arm base, inside the posture band below. Picks and places reach x 0.25 m.
4. Motion.
   • The leg is a sequence of cubic Bézier segments timed by minimum-jerk profiles: approach, transport, retreat.
   • Paths are arcs and diagonals rather than straight vertical moves; arrivals at the cube are tilted 25–55° from vertical. The peak speed of each segment is drawn separately (median 0.49 m/s before the limits below).
   • The close command leads the arrival at the cube by 1.0–1.5 times the predicted closing time, so the fingers close during the final descent; the arm lifts 0–0.08 s after arriving, while the fingers are still closing.
   • The predicted closing time runs from the open position to the cube's span across the pads (48 mm × (cos r + sin r) for a cube turned r from square with them).
   • The open command leads the release by 0–0.08 s; the arm leaves the release point 0.05–0.15 s after arriving.
   • Gripper tilt: upright, unless the grasp or release point cannot be reached upright; then the smallest lean towards or away from the arm base that reaches it, in 1° steps up to 20°, with the closing axis at most 15° out of horizontal. Tilt and yaw change along each segment with the position's profile.
   • Yaw at the grasp: from the start yaw towards the nearest of the four yaws square with the cube's faces, stopping short of it by an angle uniform in 0–20° (median 9.0°).
   • Yaw at the release: the end yaw plus a normal offset with σ 14° (median |offset| 9.4°); the held cube is not squared with the target.
   • Start and end poses: the real rig's teleoperation home pose, plus noise of 2.5 cm in x and y, 1.5 cm in z and 0.1 rad in yaw; the gripper upright.
5. Perturbation (72 episodes):
   • when: in the second half of the approach to the cube, or in the transport with the cube held; an approach too short for the push fails the attempt, so most pushes fall in the transport;
   • what: the arm decelerates along its path to rest while its pinch point is displaced by 1–5 cm horizontally (median 3.3 cm) and −1 to +2 cm vertically, over 0.2–0.5 s;
   • then: the rest of the leg (re-approach and grasp, or transport) is planned anew from the displaced pose;
   • the steps whose next state is driven by the push have command/right/source = "perturbation"; the leg's other steps are "script". sim/motion records the phase, first step, length and offset.
6. Late close and regrasp, drawn with probability 0.15 (34 episodes):
   • The arm reaches the grasp point with the open fingers around the cube, pauses 0.12–0.20 s and starts lifting towards the place side, 15–50° from vertical along the fingers' plane.
   • The close command comes too late: it is timed so that the fingers reach the cube's width only after the fingertips have cleared the cube top, so they close fully on air.
   • The arm hovers. 0.1–0.3 s after the fingers have closed, the gripper reopens; at least 0.7 s later the arm descends straight back to the grasp point and grasps the cube.
7. Limits; segments that exceed them are slowed down:
   • flange speed 0.65 m/s; turn rate 2.5 rad/s (yaw and tilt together);
   • joint speed 2.5 rad/s (the real rig's teleoperation limit); joint acceleration 10 rad/s² (the teleoperation node uses 6 rad/s²);
   • elbow bend ≥ 20°; wrist centre ≥ 0.07 m from the base axis;
   • posture band, on every solved pose: shoulder lift −130° to −29°, wrist 1 −121° to −52°, wrist 2 −112° to −67°, elbow bent at least 30° (the range of the real teleoperation, without the real episodes whose elbow passes through straight at full stretch). No episode's elbow changes sign.
8. Success check. An episode is kept only if, at its end:
   • the cube is within 2.5 cm (xy) and 1.2 cm (z) of its planned place pose and tilted less than 15°;
   • the arm's fingers are off it;
   • every other object moved less than 1.5 cm.

The episodes come from one collection run (170,077 attempts, most of them arrangements without a task for the requested cell, which are rejected before simulation). Of 357 executed attempts, 300 succeeded; the failures were 53 placed off in xy, 3 off in z and 1 with the fingers still on the cube. Grasps far from square with the cube fail more often: in a test with the grasp yaw fixed 20° short of square, 32 of 45 executed attempts succeeded, against 32 of 35 when square.

HDF5 layout

The layout of the real datasets (see their cards), with these differences:

• Not present: observations/scale, teleop/, params/, command/<arm>/eef_twist, command/<arm>/boundary_limited.
• Added: sim/, described below.

attrs: language_instruction "" (empty), episode_id, record_hz 30.0, timestamp "sim episode <id>", sim True,
       compress, cameras ["top","left_wrist","right_wrist"], arms ["left","right"], eef_frame,
       fingertip_offset_m [0, 0, 0.254]
observations/{left,right}/
    joint_pos (T,6) rad   joint_vel (T,6) rad/s
    joint_current (T,6)   actuator torque in N m (the real datasets store motor current in A)
    eef_pose (T,7)        flange (tool0) in the arm's base frame, xyz m + quaternion xyzw
    gripper_opening (T,)  0..1 of the stroke (1 = fully open; the open command is 0.433)
    gripper_state (T,)    uint8, emulating DH register 0x0201:
                          0 moving; 1 at target (open, or closed on nothing);
                          2 object caught (closing, pads on a cube, opening no longer shrinking);
                          3 object dropped (contact lost while still closing after a catch)
observations/images/{top,left_wrist,right_wrist}       (T,max_len) uint8, zero-padded JPEG bytes
observations/images/{top,left_wrist,right_wrist}_len   (T,) int32 JPEG byte count
command/{left,right}/
    source (T,) str       "script" while the arm plays its leg, "perturbation" at the steps whose next state is
                          driven by a push, "idle" before and after its leg (the left arm: always "idle")
    engaged (T,) bool     true during the arm's leg
    joint_vel (T,6)       (joint target k − joint target k−1) × 30, rad/s
    gripper (T,)          0 open / 1 closed
timestamps/ step, top, left_wrist, right_wrist, left_state, right_state
                          simulated time (s): 1.8e9 + 3600 × episode_id + step / 30
events/ time, arm, kind ("gripper"), value (1 close / 0 open), detail ("")
sim/
    object_names (5,)     cube_pink, cube_orange, cube_blue, box_white, plate
    object_pose (T,5,7)   bottom centre xyz (workcell frame, m) + quaternion wxyz, per step
    wrist_camera_noise (2,4)   orientation noise quaternion (wxyz) of the left and right wrist cameras
    top_camera_offset (2,)     pan and tilt of the top camera, degrees (pan positive turns the view to the left seen
                               from above, tilt positive raises it)
    lighting/  attrs shape, tilt_deg, size, intensity; centre (3,), colour (3,);
               per light: pos (L,3), dir (L,3), diffuse (L,3), cutoff (L,), exponent (L,), castshadow (L,)
    task/      attrs arm ("right"), place_arm ("right"), relay (False), cube, source, target_kind, target, band,
               distance; pick_position (3,), pick_yaw, place_position (3,), place_yaw: the cube's bottom centre and
               yaw before the grasp and at the planned placement
    motion/    attrs of the leg: the sampled motion parameters, among them
               miss ("none" | "late"), miss_clear, miss_dwell, lift_len (late close),
               close_lead (times the predicted closing time), grasp_dwell (s after arriving before lifting),
               speeds (4,): peak flange speeds of approach, re-approach (late close only), transport, retreat, m/s;
               arm, start_step;
               close_steps, open_steps (steps of the gripper commands),
               grasp_step, release_step (steps at which the fingers arrive at the grasp and the release);
               perturbation ("none" | "approach" | "carry"), perturbation_start (step, −1 if none),
               perturbation_steps, perturbation_offset (3,) m;
               grasp_tilt, release_tilt (2,): gripper tilt at the grasp and the release as [incline, pitch] in rad
               (incline: the closing axis turned out of horizontal; then pitch: the finger plane turned about the
               closing axis; tilt from vertical arccos(cos incline · cos pitch));
               grasp_slack (rad, the grasp yaw's shortfall from square), release_offset (rad, the release yaw
               about the end yaw), start_yaw, end_yaw (rad about the home yaw), start_offset, end_offset (3,) m

Decode a frame: cv2.imdecode(f["observations/images/top"][i, :f["observations/images/top_len"][i]], 1).

episodes.csv

One row per episode, shown in the dataset viewer.

column	meaning
file_name, episode_id	the episode file and its id
task	language_instruction; empty
steps, duration_s	recorded steps (30 Hz) and duration
arm, place_arm, relay	the arm that picks and places the cube (right), and False
cube	cube_pink, cube_orange or cube_blue
source, source_kind	what the cube rests on before the grasp (table, box_white, plate or a cube's name) and its kind
target, target_kind	where it is placed, and that target's kind
band, distance_m, distance_bin	transport band (short < 0.16 m, medium < 0.30 m, long), pick-to-place distance in xy, and its 5 cm bin
late_close	a deliberate late close and regrasp
gripper_closes	close commands of both arms
perturbation, perturbed_arm	the perturbed phase (approach, carry or none) and arm
light_shape, light_intensity	overhead light shape and mean table irradiance relative to the reference rig
light_cool_shift	(blue − red) / 2 of the light colour multiplier; positive is cooler
top_pan_deg, top_tilt_deg	the top camera's turn
grasp_tilt_deg, release_tilt_deg	the gripper's tilt from vertical at the grasp and the release
peak_flange_speed_m_s	from both arms' recorded eef_pose
pick_x/y/z, place_x/y/z	the cube's bottom centre before the grasp and at the planned placement (workcell frame, m)

Known differences from the real data

• Gripper opening while holding a cube: about 0.33 here (the 48 mm cube over a linear 0.145 m stroke). The real datasets read about 0.28 while holding a cube. The real opening reading has not been verified to be linear in the finger gap, and the real cube edge is uncertain by a few millimetres.
• Wrist cameras: the mount side matches the real rig. The rest of the mount pose and the field of view come from the model, not a calibration. The real wrist views frame the fingers differently: the fingertips appear only along the top edge, while here the gripper body fills the upper part of the image. The per-episode orientation noise does not cover that difference.
• joint_current is the simulated actuator torque in N m, not motor current.
• gripper_state is derived from contacts and the opening rate, not read from a gripper register.
• Gripper tilt: real teleoperators tilt the gripper a little at most grasps and releases (median 6–9°, p95 13–25° over the three real recordings; wrist 2 speed p95 0.32–0.62 rad/s). Here the gripper is upright except where a pose needs a tilt (wrist 2 speed p95 0.02 rad/s).
• Motion style: the scripted motions are faster and smoother than teleoperation, with joint accelerations up to 10 rad/s² where the teleoperation node is set to 6. Their only hesitations and corrections are the deliberate late closes and the perturbations.
• Rendering: rasterised with shadow maps. Shadows are harder-edged than under the real room lighting, and there are no reflections or camera noise beyond JPEG compression. The table and the backdrop have no texture. An idle arm's wrist camera gives identical frames from step to step.
• Top view background: beyond the green backdrop's side edges the real top images show a light wall in the two upper corners, the simulation the backdrop. The arm mount plates are white boxes here; on the real rig one is dark (right arm) and one clear (left arm).

Versions

revision	content
this version	300 right-arm episodes for the rig's setting of 2026-11-03: top camera at 640 × 360 (16:9) at the pose fitted to pick_and_place_quick_new revision 88bfa62, table edges re-measured from it; otherwise as af90da1
af90da1	300 right-arm episodes: the gripper tilts where an upright gripper cannot reach, the table region extends to x 0.30 m, a posture band keeps the arm away from stretched and folded configurations, yaw turns like the teleoperation (grasps not always square, releases near the home yaw), start and end yaw noise halved
98ee80d	300 right-arm episodes for the rig's setting of 2026-11-02: top camera refitted and turned ±3° per episode, backdrop colour and cooler light colours matched to pick_and_place_quick_new, the table no longer a target, faster motion, earlier closes with lifting while the fingers close, faster gripper opening
5cd8956	1000 episodes: two-arm relays for moves beyond one arm's reach, perturbed episodes, even pick-to-place distances up to 0.8 m, cells balanced over source and target kind, light shape and arm, gravity-compensated arms
7fa3c58	500 single-arm episodes, re-collected with both wrist cameras on the table side of the grippers and with step 0 rendered under the episode's own lights and wrist camera noise
bddaa31	the 93 early-close regrasp episodes of 73c6ccd replaced by late-close regrasps; left wrist camera on the robot side
73c6ccd	first upload: 500 episodes, 93 of them with an early close above the cube and a regrasp; left wrist camera on the robot side

The π0.5 checkpoints in the model repo JingwuLuo/ur3_sim_pick_and_place_cube were trained on revision 7fa3c58 (sim/, sim_real/) and 5cd8956 (sim2/, sim2_mix/).