FMB single-object (LeRobot v3)

A LeRobot Dataset v3 port of the FMB (Functional Manipulation Benchmark) single-object manipulation demonstrations, recorded with a Franka Panda arm.

This is a reformatted derivative, not the original release. The original data and full documentation are published by the authors: https://huggingface.co/datasets/charlesxu0124/functional-manipulation-benchmark Paper: arXiv:2401.08553 · Project: https://functional-manipulation-benchmark.github.io

What this is

FMB ships one .npy per demonstration (4 RGB + 4 depth cameras, proprioception, 6-axis end-effector force/torque, a commanded cartesian action, and per-step skill primitives). This port converts each single-object demonstration into one LeRobot episode, keeping the RGB streams, proprioception, force/torque, and action on a uniform frame grid.

• Episodes: 1844
• Frames: 418,495 @ 10 fps
• Robot: Franka Panda
• Cameras: side_1, side_2, wrist_1, wrist_2 (RGB 256×256)
• Per-frame task: the active skill primitive (e.g. grasp, insert, rotate)
• Scope: single-object subset only (FMB's multi-object subset is not included in this port).

Features

key	dtype	shape	notes
observation.images.{side_1,side_2,wrist_1,wrist_2}	video	256×256×3	RGB (converted from FMB's BGR)
observation.state	float32	(28,)	joint pos (7) + joint vel (7) + EE pose (7) + EE vel (6) + gripper (1)
observation.state.joint_position	float32	(7,)
observation.state.ee_pose	float32	(7,)	xyz + quaternion, base frame
observation.state.gripper	float32	(1,)	0=open, 1=closed
observation.force	float32	(3,)	end-effector force, EE frame
observation.torque	float32	(3,)	end-effector torque, EE frame
observation.jacobian	float32	(42,)	robot jacobian (6×7), flattened
action	float32	(7,)	commanded cartesian: xyz, rpy, gripper

Per-episode object metadata (shape/size/length/color/angle/distractor + object_info) is in meta/fmb_episodes.json.

Fidelity notes (please read)

• Depth dropped. FMB's 4 depth maps are not included in this port (RGB + F/T + proprio
  • action only). Use the original dataset if you need depth.
• BGR → RGB. FMB stores images in BGR; they are converted to RGB here.
• Action is the FMB commanded action as-is (no next-pose reconstruction).
• fps = 10 is nominal. The source .npy carry no timestamps; frames map 1:1, so fps is metadata, not a resampling rate.

Citation

@article{luo2024fmb,
  title   = {FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning},
  author  = {Luo, Jianlan and Xu, Charles and Liu, Fangchen and Tan, Liam and Lin, Zipeng and Wu, Jeffrey and Abbeel, Pieter and Levine, Sergey},
  journal = {arXiv preprint arXiv:2401.08553},
  year    = {2024}
}

Conversion scripts: https://github.com/lvjonok/fmb-lerobot-port