FurnitureBench

A reproducible real-world furniture-assembly benchmark from KAIST and others, testing long-horizon, precise robot manipulation.

FurnitureBench was published at RSS 2023 by KAIST's CLVR Lab in Korea and UC Berkeley, as a reproducible real-world furniture-assembly benchmark. It uses a Franka Panda arm and Intel RealSense cameras, with the task of assembling 8 kinds of furniture — a lamp, a square table, a desk, a drawer, a cabinet, a round table, a stool, and a chair — requiring multi-step coordination of grasping, insertion, and tightening, making it a long-horizon, contact-rich manipulation task. To let different labs set up a consistent environment, the parts are 3D-printed, the hardware uses common off-the-shelf products, and detailed build instructions are included; the authors also provide about 219.6 hours and 5,100 teleoperated demonstrations, plus FurnitureSim, a simulated version built on Isaac Gym and Factory. Evaluation of offline reinforcement-learning and imitation-learning algorithms shows there's still a lot of room for improvement on tasks like these.

ExampleScrewing one table leg into a tabletop requires roughly 540 degrees of rotation; limited by wrist rotation range, the robot has to perform at least 5 “rotate 90 degrees, release, and re-grasp” cycles, which is exactly what makes this benchmark hard.

Also called

Reproducible Real-World Benchmark for Long-Horizon Complex Manipulation, FurnitureSim

Related

Robotic Assembly · Long-horizon Task · Imitation Learning · Offline Reinforcement Learning · Factory / IndustReal · Isaac Gym

Sources

FurnitureBench project page
FurnitureBench (arXiv 2305.12821)

As of

2023-07