chomeed/mimicgen_threading_d0_224x224_insertcam

mimicgen_threading_d0_224x224

MimicGen Threading_D0 demonstrations in LeRobot v3 format, with both cameras rendered at 224x224. Same pipeline and conventions as chomeed/mimicgen_coffee_d1_224x224.

episodes	1000
frames	224508
fps	20
robot	Panda (OSC_POSE, control_freq 20)
cameras	agentview, eye_in_hand -- 224x224, AV1 crf 30
observation.state	37 = 9 proprioceptive + 28 privileged object dims

Provenance

Built from NVIDIA's released core/threading_d0.hdf5 (amandlek/mimicgen_datasets). That file already contains every low-dim observation, but its images are only 84x84, so only the cameras were recomputed: each frame's recorded states vector is written back into a live Threading_D0 and both views are re-rendered at 224x224. Every other column is copied verbatim from the source hdf5, so this is a re-render, not a re-simulation, and no dynamics were replayed.

Fidelity check: re-rendering demo 0 at the source's own 84x84 and diffing against its stored images gives mean |difference| of 3.10 (agentview) and 0.79 (eye_in_hand) out of 255, versus 73.5 / 91.0 for a vertically flipped comparison. The small residual is expected -- the originals were rendered on a different OpenGL stack.

Schema

Follows the conventions of chomeed/mimicgen_coffee_d0_224x224:

• observation.state (37) = concat(robot0_eef_pos, robot0_eef_quat, robot0_gripper_qpos, object), i.e. 9 proprioceptive dims followed by 28 privileged object dims. Slicing to the leading 9 gives a proprioception-only policy input.
• observation.sim_state (45) replays into a live Threading_D0.
• success == done == (reward > 0).
• source_demo_index == episode_index -- the demo's index in the source hdf5.
• is_first / is_last / is_terminal mark only the first and last frame of an episode.

Column widths are task-specific: coffee carries a 57-dim object block plus *_rel_pod keys, square a 17-dim block with no rel_* keys.

License

cc-by-4.0, inherited from the MimicGen datasets this is derived from.

Added camera: observation.images.insertion

A third 224x224 view rendered by sim-state replay, on top of the original agentview and eye_in_hand. The existing wrist camera looks straight down the gripper axis, so the tripod's bore is never visible; this one is placed to look through it, which is what makes needle-to-hole alignment readable rather than inferred.

The mount is a single fixed pose in the robot0_right_hand body frame, solved once by aiming at the needle/tripod midpoint at the closest-approach frame of episode 0 (az 260 deg, elev 15 deg, r 0.15 m) and reused unchanged for every episode of all three splits, so the three datasets share one camera.

cam_pos_local  = [-0.19815, -0.035687, 0.078038]
cam_quat_local = [-0.322727, 0.517794, 0.611714, -0.50353]   # w, x, y, z

Also in meta/insertion_camera.json. Encoded to match the existing views exactly (AV1, yuv420p, 20 fps, crf 30), and the per-episode video timestamps reproduce the source agentview values.

Known limitation: the mount is rigid to a moving hand, so for roughly the first half of each episode the tripod is out of frame. It is an insertion-phase view, and is meant to be used alongside the other two cameras, not instead of them.