RoboDyna/robodyna-benchmark-v1

RoboDyna benchmark — collection v1

Bimanual UR5-WSG manipulation trajectories collected in SAPIEN (RoboTwin 2.0 fork). Scripted-expert rollouts; only successful episodes are kept.

Contents

path	format
lerobot/<task>__<op>/	LeRobot v2.1 — parquet states/actions + mp4 per camera + meta/
hdf5/<task>/<op>/	raw episode*.hdf5 (jpeg-encoded frames + per-step obs)

Cameras: head (countertop view), left_wrist, right_wrist — all Intel D435 at 320x240.

<op> is the scenario condition: base (default), opt1, opt2 (harder dynamic variants).

Coverage

• combos with data: 78 / 82 in the target matrix
• combos at the full 50-episode target: 78
• total episodes: 4051

task	op	episodes (hdf5)	lerobot parquet	complete
boil_milk	base	50	50	yes
catch_cuboid	base	50	50	yes
catch_cuboid	opt2	50	50	yes
catch_cup	base	50	50	yes
catch_marbles_trapdoors	base	50	50	yes
catch_marbles_trapdoors	opt1	50	50	yes
catch_marbles_trapdoors	opt2	50	50	yes
catch_mouse_object_drop	base	50	50	yes
catch_ramp_ball	base	50	50	yes
catch_ramp_ball	opt1	50	50	yes
catch_ramp_ball	opt2	50	50	yes
catch_shelf_marble	base	50	50	yes
catch_shelf_marble	opt1	50	50	yes
catch_shelf_marble	opt2	50	50	yes
catch_valley_ball	base	50	50	yes
catch_valley_ball	opt1	50	50	yes
catch_valley_ball	opt2	50	50	yes
clean_table	base	50	50	yes
control_quality	base	50	50	yes
control_quality	opt1	100	100	yes
control_quality	opt2	50	50	yes
cook_meat	base	50	50	yes
cook_meat	opt1	50	50	yes
cook_meat	opt2	50	50	yes
cook_meat_timer	base	50	50	yes
cook_meat_timer	opt1	100	100	yes
cook_meat_timer	opt2	50	50	yes
dispense_gummy	base	50	50	yes
dispense_gummy	opt1	50	50	yes
dispense_gummy	opt2	50	50	yes
drop_ball_hole	base	50	50	yes
drop_ball_hole	opt1	50	50	yes
drop_ball_hole	opt2	50	50	yes
fill_coffee_jar	base	50	50	yes
hit_target	base	50	50	yes
hit_target	opt1	50	50	yes
hit_target	opt2	50	50	yes
load_train	base	50	50	yes
load_train	opt1	50	50	yes
load_train	opt2	50	50	yes
make_soup	base	50	50	yes
marble_shelf_maze	base	50	50	yes
marble_shelf_maze	opt1	50	50	yes
marble_shelf_maze	opt2	50	50	yes
measure_ingredient	base	50	50	yes
pack_fruits	base	50	50	yes
pack_fruits	opt1	50	50	yes
pack_fruits	opt2	50	50	yes
pick_ripe_apple	base	50	50	yes
pick_ripe_apple	opt1	50	50	yes
pick_ripe_apple	opt2	51	51	yes
place_block_belt	base	50	50	yes
place_block_belt	opt1	50	50	yes
place_block_belt	opt2	50	50	yes
play_billiard	base	50	50	yes
play_billiard	opt1	50	50	yes
play_billiard	opt2	50	50	yes
punch_dual_holes	base	50	50	yes
punch_dual_holes	opt1	50	50	yes
punch_dual_holes	opt2	50	50	yes
put_cup_belt	base	50	50	yes
put_cup_belt	opt1	50	50	yes
put_cup_belt	opt2	50	50	yes
save_goal	base	50	50	yes
save_goal	opt1	50	50	yes
save_goal	opt2	50	50	yes
serve_dinner	base	50	50	yes
sort_apples_belt	base	50	50	yes
sort_apples_belt	opt1	50	50	yes
sort_apples_belt	opt2	50	50	yes
stop_ball	base	50	50	yes
stop_valley_ball	base	50	50	yes
stop_valley_ball	opt1	50	50	yes
stop_valley_ball	opt2	100	100	yes
trap_bug	base	50	50	yes
whack_moles	base	50	50	yes
whack_moles	opt1	50	50	yes
whack_moles	opt2	50	50	yes

Combos with no data yet (collection still running) are omitted from the table.

Loading

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset('<local-clone>/lerobot/cook_meat__base')

Generated 2026-08-16.