[Paper Review] RoboAssembly: Learning Generalizable Furniture Assembly Policy in a Novel Multi-robot Contact-rich Simulation Environment

Mingxin Yu, Lin Shao+5 more|arXiv (Cornell University)|Dec 19, 2021

Innovations in Concrete and Construction Materials4 citations

TL;DR

This paper introduces RoboAssembly, a multi-robot, contact-rich simulation environment for learning generalizable furniture assembly policies using deep reinforcement learning. By combining structured representations with model-free RL and imitation learning, the approach achieves 74.5% success rate on unseen chairs under object-centric settings—significantly outperforming RRT-Connect baseline (18.8%) with faster planning and better generalization.

ABSTRACT

Part assembly is a typical but challenging task in robotics, where robots assemble a set of individual parts into a complete shape. In this paper, we develop a robotic assembly simulation environment for furniture assembly. We formulate the part assembly task as a concrete reinforcement learning problem and propose a pipeline for robots to learn to assemble a diverse set of chairs. Experiments show that when testing with unseen chairs, our approach achieves a success rate of 74.5% under the object-centric setting and 50.0% under the full setting. We adopt an RRT-Connect algorithm as the baseline, which only achieves a success rate of 18.8% after a significantly longer computation time. Supplemental materials and videos are available on our project webpage.

Motivation & Objective

• To develop a realistic, contact-rich multi-robot simulation environment for complex furniture assembly tasks.
• To formulate robotic furniture assembly as a concrete reinforcement learning problem with structured state representations.
• To learn a generalizable policy capable of assembling diverse, unseen chair designs using multi-robot cooperation.
• To improve upon traditional motion planning methods like RRT-Connect in both success rate and computational efficiency.
• To demonstrate zero-shot generalization of the policy to novel chair configurations not seen during training.

Proposed method

• Design a novel multi-robot simulation environment with two Franka arms on mobile platforms, supporting complex contact-rich manipulation.
• Integrate motion-planning modules to enable collision-free, coordinated movement of multiple robots during assembly.
• Construct a dataset of 220 chairs from PartNet, annotated with graspable regions and connection points for realistic assembly tasks.
• Formulate the assembly task as a hierarchical reinforcement learning problem using an object-centric and full-observation state representation.
• Train a unified policy network via imitation learning and model-free RL, leveraging data augmentation from 640 single-task policies.
• Use a multi-task training pipeline to learn a generalizable policy that generalizes to unseen chair designs in zero-shot fashion.

Experimental results

Research questions

• RQ1 Can a deep reinforcement learning policy generalize to unseen furniture designs in a multi-robot, contact-rich assembly environment?
• RQ2 How does the proposed RL-based policy compare to traditional motion planning (e.g., RRT-Connect) in terms of success rate and planning efficiency?
• RQ3 To what extent does the object-centric state representation improve policy generalization compared to full-observation settings?
• RQ4 Can imitation learning combined with model-free RL effectively transfer skills from single-task to multi-task furniture assembly?
• RQ5 Does the integration of motion planning modules into the RL pipeline enable faster and more reliable path planning than standalone planners?

Key findings

• The proposed method achieves a 74.5% success rate on unseen chairs under the object-centric setting, significantly outperforming the RRT-Connect baseline (18.8%).
• Under the full setting, the method achieves a 50.0% success rate on unseen chairs, demonstrating strong zero-shot generalization capability.
• The trained policy plans feasible assembly paths in an average of 37.8k steps per chair, compared to 91.7k steps for RRT-Connect, indicating superior computational efficiency.
• The multi-task policy trained on 640 augmented single-task policies generalizes effectively to 192 unseen chairs, confirming robustness and scalability.
• The method reduces average planning time to under two minutes per successful assembly, while RRT-Connect requires significantly longer computation.
• The success rate on the hard training set is 79.4% under object-centric setting, indicating strong performance on complex, challenging chair configurations.