FMB (Functional Manipulation Benchmark)
UC Berkeley (RAIL / Robotic AI & Learning Lab) · 2024 · datasets.bot · datasets.bot page
One-liner. Berkeley real-world Franka Panda benchmark: 22,550 expert demos of grasping, repositioning, and assembly with RGB-D, proprioception, and force/torque.
Setup
- Datasets / benchmarks: FMB is a real-world robotic manipulation benchmark from UC Berkeley's RAIL lab for studying functional manipulation, where a Franka Panda arm composes skills (grasping, repositioning, assembly/insertion) using 3D-printed objects on assembly boards. It comprises 22,550 expert demonstration trajectories across single-object and multi-object multi-stage tasks, recorded from 2 global and 2 wrist Intel RealSense D405 cameras (RGB + depth) plus proprioception and end-effector force/torque. Raw data ships as .npy dictionaries with code to convert into RLDS for training imitation-learning baselines. License: CC-BY-4.0. Download: https://functional-manipulation-benchmark.github.io/dataset/index.html.
- Hardware / simulator: Embodiment: franka_panda. Environment: lab, tabletop. Realness: physical.
Schema
trajectories -> steps -> {observation: {side_1, side_2, wrist_1, wrist_2 RGB+depth, tcp_pose, tcp_vel, tcp_force, tcp_torque, joint_pos, joint_vel, gripper}, action, primitive}
Links