state dict | action dict | video dict | annotation dict |
|---|---|---|---|
{
"joint_position": {
"original_key": "observation.state",
"start": 0,
"end": 7
},
"gripper_qpos": {
"original_key": "observation.state",
"start": 7,
"end": 8
}
} | {
"joint_position": {
"original_key": "action",
"start": 0,
"end": 7
},
"gripper_close": {
"original_key": "action",
"start": 7,
"end": 8
}
} | {
"exterior_1": {
"original_key": "observation.images.11845904_left"
},
"exterior_2": {
"original_key": "observation.images.37149196_left"
}
} | {
"human.task_description": {
"original_key": "task_index"
}
} |
pose6daug
Real-world Franka manipulation episodes with object-swap and action augmentation artifacts. 120 training episodes over 4 objects (blue_cup, green_pear, kanu, white_spray), dual ZED cameras (exo static + ego wrist-mounted).
Layout
Per-frame PNGs are packed into uncompressed tars per episode — the dataset has ~427k mask/plate frames and loose files hit Hugging Face's per-repo file recommendation and API rate limits hard.
data/<object>/<NNNN>/
masks.tar masks/{object,object_inpaint,robot,robot_sam2}/{exo,ego}/
plus preserved generations (object_inpaint_v1,
_v3dil, _v3raw, _pre_maskfix)
inpainted_frames.tar inpainted/{object,robot}/{exo,ego}/ plus preserved
plate generations (object_v1, object_pre_v3,
object_pre_maskfix)
inpainted/*.mp4 per-generation encodes (loose)
recorded_videos/ raw mp4: exo 37149196_*, ego 11845904_*
teleoperation.h5 robot states / actions
scripts/ pipeline code
calib_ws/ calibration_final.json, CALIBRATION_NOTES.md
external/swap_meshes/ metric swap meshes
external/sim_action_aug_code/ MuJoCo scene + rollout code
external/twin/ digital-twin geometry
actaug/swap_videos/ 40 object-swap mp4s (5 episodes)
actaug/rollout_videos/ 122 rollout mp4s (<ep>__<stage>__<config>__{exo,wrist})
Unpack with tar -xf masks.tar inside the episode directory.
Raw ZED SVO stereo recordings (56.5 GB) are not in this repo yet.
Calibration
Exo intrinsics fx 523.947, cx 667.921, cy 371.911 at 1280x720.
T_base_exocam t = [0.2565, 0.6050, 0.6329]; T_ee_egocam t = [-0.0839, 0.0296, 0.0261];
time_offset_s 0.27112.
Known issues
Ego object masks are unreliable on a large fraction of episodes. An audit with independent per-object detectors flagged bad frames in 27/30 blue_cup, 25/30 kanu (23/30 by visual review) and 27/30 white_spray episodes, and found that 22 of 60 delivered blue_cup + white_spray ego plates still contain the original object.
Cause: both mask-seeding routes seed the SAM2 ego track at frame 0 via a geometric transfer from the exo view. When the object is not visible in the ego view at frame 0 (it sits behind the wooden riser early in many episodes) the seed lands on a gripper finger and SAM2 tracks that instead. The acceptance gate passed on the worst cases.
green_pear was rebuilt with an independent image-domain detector (1239 bad frames across 22/30 episodes -> 47 frames across 15/30). blue_cup, kanu and white_spray are not yet rebuilt. Exo masks and exo plates are unaffected.
Also: the green_pear stem/leaf is excluded from the object mask on ~31% of frames (a background-negative sampling bug places a SAM2 negative prompt on the stalk).
See scripts/docs/realvideo/PIPELINE.md.
- Downloads last month
- 823