choijoshua16/office_multimodal_sweep

Office Multimodal Manipulation Sweep

A dual-arm (ALOHA) manipulation dataset collected in the RoboPRO / RoboTwin simulator. Every task is run over many scene seeds and four distinct planner motion modes, so the same (task, seed) is recorded four ways that differ only in how the arm moves — a controlled source of trajectory-level multimodality.

Contents

• 20 office tasks × 20 scene seeds × 4 motion modes = 1,600 episodes.
• Outcome tally: 919 mode_success · 646 mode_unrealized · 35 mode_failed.
• mode_unrealized = the motion variation could not be applied on that task/seed (e.g. a compositional task with no valid waypoint); the episode is still recorded and honestly labelled.

Motion modes

mode	waypoint injected	description
direct	none	planner default: straight grasp→place transport
high_arc	+lift	raises the transport arc
swing_left	+lateral	detours the transport left
swing_right	+lateral	detours the transport right

The waypoint is injected on the transport segment (the place transport for pick-and-place tasks, or the empty-hand grasp approach for tasks that never place, e.g. drawers) while the segment endpoints are held fixed.

Layout

runs/<task>/seed{NNNNN}/<mode>/
    episode.json              # metadata: task, seed, mode, outcome, segments, ...
    data/episode0.hdf5        # trajectory + multi-camera observations
    video/episode0.mp4        # rollout video (head camera)
index.jsonl                   # one line per episode (task, seed, mode, outcome, n_frames, ...)
summary.json                  # per-run tally + task/seed lists
office20_tasks.json           # the 20 task names
multimodal_demo.html          # self-contained viewer (outcome matrices + showcase videos/3D plots)

HDF5 schema (per episode, T = timesteps @ 50 fps)

• joint_action/vector (T, 14) — 14-DoF action (both arms + grippers)
• joint_action/{left,right}_arm (T, 6), .../{left,right}_gripper (T,)
• endpose/{left,right}_endpose (T, 7), .../{left,right}_gripper (T,)
• observation/<cam>/rgb — JPEG-encoded frames, for cameras: countertop_camera, head_camera, front_camera, left_camera, right_camera, demo_camera, demo_camera_2
• observation/<cam>/{depth, intrinsic_cv, extrinsic_cv, cam2world_gl}
• pointcloud — world-frame XYZ point cloud
• targeted_state, targeted_labels — target actor pose / progress annotations
• multimodal (group) attrs: mode, outcome, segments, pair_id, n_injected_waypoints

Tasks

close_drawer, open_drawer, put_book_in_fileholder, put_book_on_book, put_milktea_next_to_laptop, put_milktea_on_shelf, put_mouse_next_to_stapler, put_mouse_on_pad, put_phone_next_to_cube, put_phone_on_holder, put_rubikscube_in_drawer, put_rubikscube_next_to_milktea, put_stapler_in_drawer, put_stapler_next_to_mouse, put_stapler_on_book, store_stapler_in_drawer, store_rubikscube_on_shelf, move_items_around, organize_table, set_up_table.

Notes

• One episode (put_rubikscube_in_drawer / seed 1 / high_arc) is a 0-frame mode_unrealized case: it has episode.json but no HDF5/video.
• Language instructions per task are sampled from the RoboTwin instruction bank (not stored in the HDF5; episode.json.instruction may be null for scripted-expert runs).