FurnitureBench

FurnitureBench is a real-world furniture assembly benchmark introduced at RSS 2023, designed to push robot manipulation beyond simple pick-and-place tasks into complex, long-horizon assembly scenarios. Developed by researchers at Columbia University, it addresses the gap between simple manipulation benchmarks and the dexterity required for real-world tasks like furniture assembly. The benchmark features IKEA-style furniture assembly tasks that require multiple steps, precise alignment, force-sensitive insertion, and sequential reasoning. Unlike prior benchmarks that focus on picking or pushing, FurnitureBench requires robots to manipulate multiple parts, align holes and pegs, insert connectors, and tighten screws — all in a real physical setup with real-world physics, friction, and tolerances. FurnitureBench supports multiple learning paradigms: reinforcement learning, imitation learning, and task-and-motion planning (TAMP). It provides a standardized evaluation protocol across multiple furniture types (chairs, tables, cabinets) with varying difficulty levels, enabling systematic comparison of different approaches to long-horizon manipulation. The benchmark is particularly challenging because it tests: (1) long-horizon reasoning over 10-20+ sequential steps, (2) precise geometric alignment under uncertainty, (3) contact-rich manipulation with force feedback, (4) recovery from failure mid-assembly, and (5) generalization across furniture variants. These challenges make it a meaningful testbed for advancing robot manipulation capabilities. FurnitureBench fills an important gap in the embodied AI evaluation landscape by providing a real-world physical benchmark with standardized hardware and protocols. It complements simulation-based benchmarks by testing whether methods that work in simulation transfer to real hardware with all its imperfections. The open-source hardware designs and evaluation code enable reproducible research across labs.