moonbot_block_stack_sep22

LeRobot v3 dataset converted from ROS 2 (Jazzy) rosbag2/MCAP teleoperation recordings of a Moonbot arm stacking coloured blocks.

⚠️ Differs from gdiazsrl/three_blocks_stack_sep22

This is an independent conversion of the same source recordings, produced as a cross-check. It is not a drop-in replacement:

	this dataset	gdiazsrl/three_blocks_stack_sep22
observation.state	float32 [14]	float32 [21]
state sources	joint_read (8) + F_ee (6)	joint_read (8) + tip_pos (7) + F_ee (6)
episodes	293	296
frames	514,644	519,684
is_first/is_last/is_terminal	absent	present

It was converted with the moonbot1.yaml contract from the gdiazh/rosetta main branch, which omits /leg302/tip_pos. Gustavo's newer contract adds it. For training and deployment against the current contract, use his dataset — a policy trained on 14-dim state will not load against a 21-dim contract.

Contents

• 293 episodes, 514,644 frames, 9 tasks, 30 fps
• 3 RGB streams at 480x640: observation.images.front / .eef / .right
• action: float32 [8] — /leg302/joint_set_teleop

Excluded episodes (7 of 300)

Defective in the source, not lost in conversion:

Episode	Defect
1789028768_171476840	empty directory (0 objects)
1789027163_700200080	metadata.yaml present, no .mcap
1789028630_884682178	metadata.yaml present, no .mcap
1789097109_781118392	0-byte metadata.yaml
1789108382_783132076	0-byte metadata.yaml
1789372813_089231014	0-byte metadata.yaml
1789438114_570389747	0-byte metadata.yaml

The four 0-byte-metadata episodes hold 7.95 GiB of intact .mcap data — rosbag2 writes metadata.yaml on stop, so those recordings were interrupted. They are recoverable by regenerating metadata, and likely account for most of the 296 vs 293 gap.

Provenance

Source 1.283 TiB of MCAP -> 8.47 GiB (155x), converted with gdiazh/rosetta bag_to_lerobot.py. Every episode was hash-verified against the source before the raw bag was deleted. See MANIFEST.txt for checksums.