MIT researchers devised a method for quickly planning physically plausible assembly motions and sequences for actual assemblies.

Assembly planning is at the heart of modern industrial manufacturing's product assembly, maintenance, and recycling automation. Yet, despite its importance and extensive history of research, mechanical assembly planning remains a complex subject when given the finished state. It is because dealing with arbitrary 3D shapes is difficult, and real-world assemblies require minimal motion.

Objective

Researchers from MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), Autodesk Research, and Texas A&M University developed a method for automatically assembling accurate, efficient, and generalizable products to a wide range of complex real-world assemblies to alleviate some of these burdens. Their system establishes the order of multipart assembly and then searches for a physically plausible motion path for each step.

To validate their strategy, the team created a Spartan-level large-scale dataset comprising thousands of physically realistic industrial assemblies and motions. The proposed method can solve almost all of them, surpassing earlier methods by a wide margin on rotational assemblies, such as screws and puzzles. It's also a speed demon, solving 80-part assemblies in minutes. For example, if the goal is to assemble a screw attached to a rod, the algorithm will determine the best assembly method through two stages: disassembly and assembly. 

The algorithm applies varying forces to the screw and observes its movement using physics-based simulation. As a result, a torque spinning along the rod's central axis drives the screw to the end of the rod, where it is separated by a straight force going away from it. The algorithm reverses the disassembly path to obtain an assembly solution from individual parts during the assembly stage.

Proposed work

The researchers suggest a unique strategy for quickly planning physically plausible assembly movements and sequences for real-world assemblies in this paper. They use the assembly-by-disassembly technique and physics-based simulation to explore a smaller search space efficiently. The researchers define a large-scale dataset consisting of thousands of physically correct industrial assemblies with a range of assembly motions required to evaluate the universality of their method.

Their trials on this new benchmark show that they attain the highest computing efficiency and a state-of-the-art success rate compared to other baseline algorithms. Their method applies to rotational assemblies (such as screws and puzzles) and can solve 80-part assemblies in minutes.

Conclusion

According to the researchers, several major extensions would make assembly planning more broad and efficient. For starters, their assembly-by-disassembly technique restricts the assemblies to being just stiff. On the other hand, geometric-based techniques cannot handle deformable objects since they cannot simulate physical deformation. As a result, they believe it is worthwhile to investigate further the path of physics-based planning for generalization to deformable assemblies such as snap-fit assemblies. Second, it is critical to use geometry information in assembly planning in addition to physical feedback.

Unlike their method and baseline approaches, humans can rapidly deduce probable disassembly sequences and motions through vision, avoiding wasting time trying blocked portions or pushing parts towards dead ends. Finally, one exciting future effort is adding robotic arms to their simulation to manipulate assemblies following the predetermined path given by their approach to aid research in robotic assembly. Extending the capability of executing complex assembly autonomously and flexibly on real robots is promising.