pine_wm_real2sim_v0

Simulated UR7e + Robotiq 2F-85 demonstrations of six real Pine WM workcell tasks, rebuilt real2sim: task objects were measured from the real task videos with the calibrated D435 and generated with USDCraft v2; demonstrations were collected with cuMotion in Isaac Lab Arena (data engine). LeRobot v2.1, same layout as agibot_arena_v0 (one folder per task).

Folder	Instruction	Episodes	Frames
stack_cubes	Stack the green block on top of the red block.	50	21,639
ring_toss	Put the yellow ring onto the middle peg of the ring stand.	50	25,206
pour_tea	Pick up the teapot by its handle, pour tea into the cup, then put the teapot back.	50	39,488
apples_green_then_red	Pick up the green apple and the red apple, then place them onto the plate.	25	28,049
apples_red_then_green	Pick up the red apple and the green apple, then place them onto the plate.	25	29,214
erase_whiteboard	Erase the whiteboard	40	55,908

The robot prompt is English in every task (meta/tasks.jsonl and annotation.human.action.task_description).

Source recordings and shape variants

Each task rebuilds one real Pine WM UR7e recording. Only the scene was taken from it: object sizes, shapes, colours and the layout were measured from a single third-person (D435) frame. The simulated trajectories are new cuMotion demonstrations, not replays of the real robot motion.

Folder	Real source (layout and object measurement)
stack_cubes	Pine WM UR7e recording episode_000402 (third-person video)
ring_toss	Pine WM UR7e recording episode_000201 (third-person video)
pour_tea	Pine WM UR7e recording episode_000203 (third-person video)
apples_green_then_red	pine_wm_real/apples_green_then_red (25 real episodes), episode 0, first frame of observation.images.third
apples_red_then_green	pine_wm_real/apples_red_then_green (25 real episodes), episode 0, first frame of observation.images.third
erase_whiteboard	pine_wm_real/erase_whiteboard (48 real episodes), episode 0, first frame of observation.images.third

Shape variant 0 is the object rebuilt from the real recording; variants 1 and 2 are USDCraft alternatives with the same functional interface (same size class, grasp and goal), placed at the same measured layout.

Folder	Episodes	Shape variant	Assets (scene object → USDCraft asset ID)
stack_cubes	0–29	0	cube_green r2s_cube_green, cube_red r2s_cube_red
stack_cubes	30–39	1	r2s_cube_green_alt1, r2s_cube_red_alt1
stack_cubes	40–49	2	r2s_cube_green_alt2, r2s_cube_red_alt2
ring_toss	0–29	0	ring_board r2s_ring_board, rings r2s_ring_{green,red,yellow}
ring_toss	30–39	1	r2s_ring_board_alt1, r2s_ring_{green,red,yellow}_alt1
ring_toss	40–49	2	r2s_ring_board_alt2, r2s_ring_{green,red,yellow}_alt2
pour_tea	0–29	0	teapot r2s_teapot, cup r2s_cup
pour_tea	30–39	1	r2s_teapot_alt1, r2s_cup_alt1
pour_tea	40–49	2	r2s_teapot_alt2, r2s_cup_alt2
apples_green_then_red	0–24	0	r2s_pine_apple_green, r2s_pine_apple_red, r2s_pine_plate, r2s_pine_orange, r2s_pine_mango, r2s_pine_lemon, r2s_pine_pear
apples_red_then_green	0–24	0	r2s_pine_apple_red, r2s_pine_apple_green, r2s_pine_plate, r2s_pine_orange, r2s_pine_lemon
erase_whiteboard	0–39	0	r2s_pine_whiteboard (fixed), r2s_pine_eraser

The apple scenes keep the non-target fruit of the real recordings (orange, lemon, and in green-then-red also mango and pear) in place as scene objects. Per-episode variant and assets are also in meta/episode_provenance.jsonl.

Features

Key	Shape	Meaning
observation.state	7	Joint positions: shoulder_pan, shoulder_lift, elbow, wrist_1, wrist_2, wrist_3 (rad), finger_joint (rad)
action	7	Absolute joint position targets, same order
observation.eef_9d / action.eef_9d	9	TCP pose in the UR base frame: xyz + first two rotation-matrix columns (FK of state / action)
observation.images.ego_view	480×640×3	Calibrated third-person RealSense D435
observation.images.wrist_a_view, wrist_b_view	480×640×3	The two D405 wrist cameras on the tool flange
timestamp, frame_index, episode_index, index, task_index, annotation.human.action.task_description, next.reward, next.done, observation.img_state_delta		As in agibot_arena_v0

15 fps. Observation alignment is pre-step: observation[t] and image[t] are the state before action[t]. meta/modality.json groups arm [0:6], gripper [6:7] and eef_9d. meta/episode_provenance.jsonl gives per episode the shape variant, asset IDs, object layout, render randomization, success metrics and the source simulation demo.

Differences from agibot_arena_v0: single 6-DoF arm (7-D state/action, 9-D eef), wrist views named wrist_a_view/wrist_b_view (both on the same wrist, not left/right arms), native 640×480 video.

Randomization

• Layout (per episode, at collection): every object translated ±5 cm around its measured real pose and turned (cubes ±45°, rings ±180°, ring stand ±15°, teapot/cup ±30°), with ≥2 cm footprint clearance inside the reachable workspace. Apples tasks: round fruit turned freely, elongated fruit ±30°, ellipse footprints; the crowded green-then-red scene uses ±3 cm and ≥5 mm clearance. Erase: whiteboard ±2 cm/±5°, eraser ±5 cm/±30°. Every episode's layout is distinct.
• Appearance (per episode, at rendering, physics untouched): HDRI or uniform dome, key/fill light intensity, colour and direction, table materials, 0–3 primitive and 0–4 real-object distractors kept clear of every object and of the end-effector path over the whole episode.

Generation

States/actions were recorded first (no cameras) with cuMotion planning and a PhysX contact audit that rejects any robot contact other than the finger pads on the manipulated object; only successful episodes are included (task success checked from simulator state). Images were rendered afterwards by bounded state replay of the recorded states, validated per frame (arm joints < 1e-3 rad, all joints < 5e-3 rad, objects < 0.5 mm; the Robotiq passive finger joints are replayed on their mimic relation, which may differ from the recorded values by < 15 mrad — up to ~8 mrad in the apple grips). Per variant the first 30/10/10 (apples 25, erase 40) successful episodes that passed replay validation were selected; each trajectory appears in exactly one episode. The pour task grasps the teapot at the handle side (one finger inside the mouth, one on the ear handle), as on the real robot. The apples are placed into the plate in the instructed order. In erase_whiteboard the marker stroke is a visual-only decal of 43 segments: a segment disappears once the eraser's felt face covers it while pressed on the board; the re-render recomputes this from the recorded eraser poses and matched the collection-time erased mask exactly for all 40 episodes (success requires ≥95 % erased, here 100 %).

Known limits: simulation only, not reviewed for training eligibility yet; no liquid is simulated in pour_tea; the whiteboard stroke has no physics; FK of eef_9d agrees with the simulator TCP to ~3 mm.