Driving Robotic Assembly Using CAD Data

Industrial assembly is still one of the most manual tasks in manufacturing. A primary reason is the lack of flexibility in automated systems. Fixed automation is designed to manufacture one product but cannot be repurposed easily when the product changes. In some industries, product lines change every year (or even faster) making changes in automation prohibitively expensive. Automation is also hampered by the inability to adjust (or adapt) to variations in parts or the process. Traditional automation systems, for example, struggle when parts are presented in unstructured piles inside a bin versus being neatly lined up on a fixture. This lack of adaptivity makes some tasks hard to automate.

In recent research [1] published at the IEEE/RSJ IROS conference in Kyoto, Japan in October 2022, we’ve been examining the use of CAD data to help overcome some of these obstacles. CAD data has traditionally been used in robotics to guide collision checking and visualization. However, CAD data of the parts and product assemblies themselves are a rich source of information. We use CAD data in multiple ways: in an interactive tool to specify high-level input for the assembly process, to train a perception model capable of picking and re-orienting random parts, and to simulate and validate the different parts of our assembly plans.

We demonstrate this on a challenging task where a pair of robots work together to build a 3D lattice structure from Yinan blocks [2] (small 3D-printed components named after their creator). The blocks are initially arranged in a random pile on a table. As the robots pick the blocks, they often end up misaligned in the robot’s hand. Robot grippers are not very dexterous as compared to human hands so cannot manipulate parts in-hand like we can. Using onboard cameras, each robot helps the other figure out where the part is in their hand. They then hand off the part to each other to orient them the right way before inserting them into the assembly. An important note here is that one of the robots is not very accurate – it doesn’t know where it is very well. Using the onboard cameras though, it can still make the small adjustments to make all this work. Watch the movie, it’s magical!

So, what does our work intend to contribute? Autodesk’s software products are used every day by millions of people to design new product assemblies. Our products are also used to model, simulate, and design the make or manufacturing processes to bring those product assemblies to life. However, there is a noticeable gap between the design and make parts of this workflow – manufacturing constraints are not always incorporated into the design process from the beginning while the make process is very manual and often siloed from the design phase. Our research is targeted at bridging part of this gap by giving designers the ability to program robotic workcells capable of assembling their designs. We’re excited about what we’ve got going for this year so look for more as our research progresses!

Sachin Chitta is a Director of Research Science at Autodesk.

[1] Yotto Koga, Heather Kerrick, Sachin Chitta, On CAD Informed Adaptive Robotic Assembly, in IEEE/RSJ International Conference on Intelligent Robots and Systems, Kyoto, Japan, Oct. 2022
[2] Y. Zhang, Y. Koga, D. Balkcom. Interlocking Block Assembly With Robots. in IEEE Trans. on Automation Science and Eng., vol. 18, issue 3, 2021.