Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning (2506.05168v1)

Abstract: Multi-part assembly poses significant challenges for robots to execute long-horizon, contact-rich manipulation with generalization across complex geometries. We present Fabrica, a dual-arm robotic system capable of end-to-end planning and control for autonomous assembly of general multi-part objects. For planning over long horizons, we develop hierarchies of precedence, sequence, grasp, and motion planning with automated fixture generation, enabling general multi-step assembly on any dual-arm robots. The planner is made efficient through a parallelizable design and is optimized for downstream control stability. For contact-rich assembly steps, we propose a lightweight reinforcement learning framework that trains generalist policies across object geometries, assembly directions, and grasp poses, guided by equivariance and residual actions obtained from the plan. These policies transfer zero-shot to the real world and achieve 80% successful steps. For systematic evaluation, we propose a benchmark suite of multi-part assemblies resembling industrial and daily objects across diverse categories and geometries. By integrating efficient global planning and robust local control, we showcase the first system to achieve complete and generalizable real-world multi-part assembly without domain knowledge or human demonstrations. Project website: http://fabrica.csail.mit.edu/

• The paper presents Fabrica, a system that integrates hierarchical planning with reinforcement learning to perform autonomous multi-part assembly tasks.
• It details a method that decomposes complex assembly into subproblems such as grasp filtering, sequence planning, and motion coordination across dual arms.
• Benchmarks demonstrate an 80% zero-shot success rate in assembling objects with 5 to 9 parts, underscoring its robust performance in varied real-world settings.

Overview of "Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning"

The paper "Fabrica: Dual-Arm Assembly of General Multi-Part Objects via Integrated Planning and Learning" presents a novel approach to autonomous robotic assembly tasks, focusing on a dual-arm robotic system capable of handling complex, multi-part assembly. The authors introduce Fabrica, a system that integrates advanced planning algorithms and reinforcement learning techniques to enable robust assembly processes in real-world scenarios without relying on domain-specific knowledge or human demonstrations.

Core Contributions

Algorithms: The paper introduces a hierarchical planning algorithm that efficiently manages precedence, grasping, and motion planning for dual-arm assembly robots. It addresses the hybrid optimization problem by decomposing it into manageable subproblems, leveraging parallelization to enhance computational efficiency. For contact-rich tasks, Fabrica utilizes a lightweight reinforcement learning framework to develop generalist policies that operate effectively across diverse geometries and configurations.

Systems: The authors provide an end-to-end system capable of executing complete multi-part assemblies autonomously. This system maps CAD models to robotic execution, covering phases from automatic fixture design to sequence planning, motion grasping, and part insertion. The approach demonstrates adaptability across different dual-arm robot setups, attesting to its broad applicability.

Benchmarks: The paper offers a benchmark test suite comprising 7 multi-part assemblies with varying complexities ranging from 5 to 9 parts. These benchmarks are designed to emulate real-world assembly tasks with objects resembling industrial and everyday items. Through rigorous simulation and real-world experiments, the paper validates the system's effectiveness and generalizability.

Technical Insights

Fabrica's planning approach structures the complex assembly task by identifying precedence hierarchies, filtering feasible grasps, optimizing sequence and grasp configurations, and generating supporting fixtures. The system effectively balances meticulous offline planning with reactive online control using reinforcement learning. This synergy is evident in the successful zero-shot transfer of simulated policies to real-world applications, achieving an 80% success rate in assembly steps without external input.

Implications and Speculative Developments

The practical implications of this research are notable for industries requiring flexible automation solutions capable of adapting to new tasks with minimal downtime. Theoretically, the paper contributes to the understanding of integrated planning and learning for robotic manipulation, suggesting pathways for further exploration into multi-arm coordination, dexterous manipulation, and dynamic adaptation.

Looking forward, the paper sets a foundation for future exploration into more dexterous systems involving in-hand manipulation or reorientation during assembly processes. Addressing constraints like heavier part handling, variable grasp strategies, or integrating vision feedback for dynamic obstacle avoidance could significantly enhance the usability and scope of Fabrica-like systems.

This contribution aligns with ongoing innovations in AI-driven robotics, potentially inspiring developments in general-purpose robotic systems capable of performing intricate tasks independently, thereby broadening the scope of automated solutions in manufacturing and beyond.