UWM-test

A held-out test set sampled from four RoboMIND 2.0 robot-manipulation subsets (Franka, Tianyi, Tienkung, UR5). For every task that belongs to the test split (is_training = 0), 25 episodes are sampled. UR5 uses a pre-existing 8 × 25 subtest selection.

Total: 600 episodes, 24 tasks, ~288 GiB.

Directory layout

UWM-test/
├── uwm_test_index.csv          # master index (paths point into this folder)
├── README.md
├── franka/<task>/<episode_id>/
│   ├── trajectory.hdf5         # RGB (+ depth) trajectory
│   ├── seg.npz                 # segmentation
│   └── flow.npz                # optical flow
├── tianyi/<task>/<episode_id>/
│   ├── trajectory.hdf5         # RGB trajectory
│   ├── depth.npz               # depth (separate file)
│   ├── seg.npz
│   └── flow.b2
├── tienkung/<task>/<episode_id>/
│   ├── trajectory.hdf5         # RGB (+ depth) trajectory
│   ├── seg.npz
│   └── flow.b2
└── ur5/<task>/<episode_id>/
    ├── trajectory.hdf5         # RGB (+ depth) trajectory
    ├── seg.npz
    └── flow.b2

Notes on modalities:

• Franka / Tienkung / UR5: video_path and depth_path reference the same trajectory.hdf5 (RGB and depth are stored together), so only one HDF5 is stored per episode and both columns point to it.
• Tianyi: depth is a separate depth.npz file.
• Optical-flow extension varies by subset: .npz (Franka) or .b2 (Tianyi / Tienkung / UR5).

index file — uwm_test_index.csv

Columns:

column	description
dataset	one of franka, tianyi, tienkung, ur5
episode_id	episode identifier
task	task name
video_path	absolute path to trajectory.hdf5 under /mnt/UWM-test
depth_path	path to depth source (== video_path except Tianyi)
seg_path	path to segmentation file
flow_path	path to optical-flow file
fps	frames per second
h, w	frame height / width
n_frames	number of frames
caption	natural-language task description
is_training	always 0 (test split)

All path columns point into /mnt/UWM-test/....

Data distribution

Summary

dataset	tasks	episodes	size
franka	11	275	130 GiB
tianyi	2	50	28 GiB
tienkung	3	75	54 GiB
ur5	8	200	77 GiB
total	24	600	~288 GiB

Every task contributes exactly 25 episodes.

Tasks per dataset (25 episodes each)

franka (11 tasks)

• close_laptop_with_both_arms_from_behind
• pass_towel_and_hang_on_rack
• place_bearing_gear_from_tray_to_pile
• place_fruits_and_box_with_both_arms
• place_lemon_and_lettuce_in_plates_1
• place_mouse_and_phone_with_arms_1
• place_red_cookie_box_on_tray
• place_sandwich_from_basket_to_tray
• plug_charger_into_power_strip
• transfer_potato_from_pot_to_plate
• turn_and_place_switch_with_both_arms

tianyi (2 tasks)

• place_white_toy_into_blue_box_with_left_hand
• place_yellow_pepper_on_plate_with_right_hand

tienkung (3 tasks)

• flip_yellow_button_and_stand_it_with_both_arms
• pour_wine_from_measuring_cup_to_shaker
• press_green_start_button

ur5 (8 tasks)

• collect_tennis_balls_into_bucket
• flip_cup
• insert_tennis_ball_vertically_into_holder_with_arms
• open_upper_drawer
• place_bread_on_blue_plate_with_arms
• place_plate_on_table
• place_tea_on_tray
• setup_study_scene_with_laptop_mouse_and_lamp

Source

Sampled from RoboMIND 2.0 (Franka / Tianyi / Tienkung / UR5) test splits used by the UWM project.