zijian2022/fanuc_er4ia_stack_two_task_d4_fixed4_dr

FANUC ER-4iA: stack one cube on the other (D4: two tasks, four fixed layouts + appearance randomisation)

Simulated demonstrations for the FANUC ER-4iA cell (Rockwell-NVIDIA Fall 2026, Dataset 3), recorded in the v0.5 Isaac Sim twin of the cell. LeRobotDataset v3.0, 200 episodes, 33400 frames, 10 fps. Instructions: "Put the red cube on the blue cube." and "Put the blue cube on the red cube."

This variant. Two instructions. Episodes 0-99 are D3 unchanged ("Put the red cube on the blue cube.", task_index 0); episodes 100-199 are new, "Put the blue cube on the red cube." (task_index 1): the expert picks the blue cube and sets it on the red one, at the same four fixed layouts, 25 episodes each, with appearance draws offset from D3's (seed + 500). Both halves start from identical-looking scenes, so a policy must read the instruction (or a task id) to tell them apart; plain ACT ignores it. Recording logs: meta/record_log.json (D3 half) and meta/record_log_blue_on_red.json.

Related datasets (base, D1, D2 share 50 random layouts, seeds 0-49; D3 uses four fixed ones):

• zijian2022/fanuc_er4ia_stack_red_on_blue
• zijian2022/fanuc_er4ia_stack_red_on_blue_d1_fixwrist
• zijian2022/fanuc_er4ia_stack_red_on_blue_d2_fixwrist_dr
• zijian2022/fanuc_er4ia_stack_red_on_blue_d3_fixed4_dr

Scene and expert

• Two 10 mm cubes on the table at one of four fixed layouts (robot base frame, x m, y m, yaw deg):

Layout	Red	Blue
0	(0.385, -0.075, -15)	(0.385, 0.045, 10)
1	(0.415, -0.075, 10)	(0.415, 0.075, -15)
2	(0.385, -0.045, 15)	(0.415, 0.045, -10)
3	(0.415, -0.045, -10)	(0.385, 0.075, 15)

Episodes 0-24 are layout 0, 25-49 layout 1, 50-74 layout 2, 75-99 layout 3 (D4: the same again for episodes 100-199, the blue-on-red task).

• A scripted expert with ground-truth cube poses: approach, grasp the top cube (red; D4 blue-on-red: blue) across its faces, carry it, turn it to the other cube's yaw, set it down on top, release, lift. Only successful episodes are kept (top cube resting on the other within 5 mm).
• The arm tracks its targets exactly (no controller lag); the real cell's arm trails by ~0.8 s.

Features

Key	Shape	Content
observation.images.table/side/wrist	240x320x3	the cell's cameras, cropped and resized (below)
observation.state	7	J1..J6 deg (v5 twin convention, sim J3 = controller J3 + J2), gripper 0/1
observation.tcp_pose	6	fingertip x, y, z mm, w, p, r deg in UF0
action	7	the next joint target, the cell's "joints" action format

Camera preprocessing

Each frame is rendered at the camera's native 16:9 size, resized to 1920x1080 (as the cell serves a 1920x1080 request), cropped to a fixed 4:3 window and resized to 320x240. Table and side windows are the image of the box x 0.34-0.46 m, y -0.11-0.11 m, z 0-0.08 m above the surface under the cells' calibration; the wrist window is central. A policy must apply the same crop to the cell's 1920x1080 frames (act_stack_io.preprocess).

Camera	Window (x0, y0, x1, y1) in 1920x1080	Size
table	(818, 621, 1151, 871)	333x250
side	(849, 627, 1210, 898)	361x271
wrist	(420, 135, 1500, 945)	1080x810

Known limitations

• Rendered images: colours, lighting and background differ from the real cell (D2 and D3 randomise some of this).
• The twin's wrist camera sits about 6 deg (~95 px) from where the real one images the table.
• Demonstrations are perfectly consistent (no recovery from off-nominal states).

Load with LeRobotDataset("zijian2022/fanuc_er4ia_stack_two_task_d4_fixed4_dr"); pass video_backend="pyav" if torchcodec has no system FFmpeg.