play_teleop_0904

Long-horizon, unscripted "play" teleoperation episodes on a Franka (DROID setup), collected 2026-09-04 to 2026-09-07 with scripts/tests/john_play_teleop/collect_play.py (Oculus VR controller, right hand). No task instruction; the operator freely manipulates objects on the table.

episodes	total steps	total duration	control rate
25 (0000–0024)	56,516	~99.2 min	10 Hz nominal (~9.5 Hz effective)

ep	steps	sec	ep	steps	sec	ep	steps	sec
0000	2814	298	0009	1429	151	0018	1974	208
0001	4539	481	0010	2439	257	0019	1902	200
0002	3243	344	0011	303	32	0020	5121	540
0003	2824	300	0012	321	34	0021	2190	231
0004	2552	271	0013	1684	178	0022	3929	414
0005	2878	305	0014	563	59	0023	677	71
0006	1872	198	0015	558	59	0024	4721	497
0007	2746	289	0016	1038	109
0008	3000	300	0017	1199	126

Episodes 0011 and 0012 are short (~30 s). Episode 0008 has no num_steps/duration_sec/success attrs (writer was not closed cleanly); its data is complete (3000 steps). Episodes 0013–0024 were collected 2026-09-07.

Layout

NNNN/
  teleoperation.h5           # observation/*, action/*  (no images inside)
  videos/index.json          # fps, segment_frames, per-camera segment list
  videos/<cam_id>_{left,right}/seg_0000.mp4, seg_0001.mp4, ...   # 3 ZED cameras x 2 eyes, 10 fps

Timestep i is frame i % segment_frames of segment i // segment_frames (segment_frames=3000, also in h5 attrs). Cameras: 11845904, 34022131, 37149196 (ZED serials).

h5 contents (per step, length = num_steps)

• action/: cartesian_position (6), cartesian_velocity (6), joint_position (7), joint_velocity (7), gripper_position, gripper_velocity, target_cartesian_position (6), target_gripper_position, robot_state/*
• observation/robot_state/: cartesian_position (6), joint_positions (7), joint_velocities (7), gripper_position, joint/motor torques, controller latency
• observation/camera_intrinsics/<cam>_{left,right} (3x3), observation/camera_type/<cam>
• observation/controller_info/: controller_on, movement_enabled, success, failure
• observation/timestamp/: camera, control-loop and robot-state timestamps (ns)
• attrs: episode, num_steps, duration_sec, hand, stop_reason, success, task, video_fps, video_segment_frames

Join video segments into one mp4 per camera with python scripts/tests/john_play_teleop/merge_video_segments.py <episode_dir> (in the droid repo).