Evaluating Robot Performance in Assembly

The effectiveness of robots in assembly tasks varies significantly. On one hand, traditional industrial robots, which perform tasks such as picking, placing parts, and spot welding, hold a dominant market share of about 97%. These robots operate on predefined paths and are the backbone of major manufacturers like ABB, Fanuc, and Kuka. Conversely, newer "collaborative" robot arms, designed to work alongside humans and equipped with increasing autonomy through vision and tactile feedback, are struggling to gain traction.

The challenge of "greenfield" versus "brownfield" automation has long plagued providers of automation solutions. Constructing a well-balanced assembly line from the ground up (greenfield) is considerably easier than retrofitting existing environments (brownfield). To illustrate, consider the process of automating manual coffee brewing versus using a single-button espresso machine. While retrofitting can be achieved, it rarely matches the efficiency of a purpose-built solution.

Can We Assemble Intricate 3D Structures?

With this foundational understanding of assembly and the historical role of robotics, we can now explore the classification of current robotic assembly solutions and envision their future development.

The following video showcases the bimanual assembly of a mechatronic widget, a challenge addressed during the 2018 World Robot Summit Industrial Assembly Challenge. This process requires three distinct tools, precise insertions, and dexterity to hold various components while securing screws.

Preparing parts and tools for assembly, known as "kitting," is a task in itself and typically involves organizing bulk materials onto a "kitting tray" for transport to the assembly station. My company's solution, which received top marks in this category at the World Robot Summit, is depicted in the video below.

Instead of merely streamlining the overall assembly sequence, the WRS focused on optimizing the work at individual stations, embodying the principle of "Level 5 Automation." This concept is akin to the "Level 5 Autonomy" framework in autonomous driving, aiming for seamless transitions in production without the need for custom tooling and enhanced learning for continuous operation.

The WRS 2018 competition emphasized "Level 4" automation and included unexpected tasks, such as replacing pulleys and rubber belts with gears and chains. This prompted a wide array of solutions, from custom machinery to two-arm manipulation approaches that minimized jigging.

The video below presents the JAKS team's second-place solution, featuring two industrial arms equipped with specialized end-effectors for grasping and securing parts.

The performance data from all participating teams shed light on the kinds of tasks successfully accomplished, including the "Tasboard" assembly task, which involves simplified 2D assembly of all components from the 3D assembly. Notably, many teams excelled in the "Kitting" task, with the winning team, SDU, leveraging numerous 3D-printed jigs and specialized tools.

The Importance of Force and Torque Control

A key takeaway from the competition was the necessity of controlling not just the robot's position, but also the forces and torques exerted at the end effector. Historically, industrial robots lacked this capability, which has only recently been integrated into collaborative robots, exemplified by the Franka Emika Panda.

The process of peg-in-hole assembly highlights this need. Using a force-torque sensor at the robot's wrist enables the measurement of forces in three dimensions. An algorithm could employ this data to guide the workpiece into the assembly location while compensating for any misalignments.

Innovative Solutions for Robotic Assembly

Despite Fanuc's success in assembling a complete 3D structure, many solutions remain far from achieving "Level 5" automation. In a rematch held in 2020, the competition maintained its original goals but shifted focus towards enhancing autonomy. To facilitate this, an autonomous ground vehicle (AGV) was introduced to transport kitting trays to assembly and inspection stations.

Recent technological advancements have simplified several assembly tasks. For instance, Bosch/Rexroth has created an adapter that allows stiff industrial robots to exhibit compliance, facilitating tighter fits through controlled movements. Automatic screwdrivers that integrate seamlessly with collaborative robots have also made screwing operations more efficient.

The Rise of Humanoid Robots

Recent developments in humanoid robotics are unfolding at an astonishing rate. Companies like Figure have secured substantial funding and initiated pilots with major automotive manufacturers, while Nvidia is creating a foundational model for humanoid robots. Notably, a humanoid robot capable of executing a backflip is available for purchase, albeit without hands.

Humanoid robots may prove advantageous for several reasons:

• They can seamlessly integrate into existing human-centric processes, addressing the "brownfield" challenge.
• The collaborative robotics market has identified numerous applications where humanoids could play a significant role.
• The advancements in AI during 2023 could lead to a transformation akin to the impact of railways and the internet.

AI Breakthroughs Supporting the Humanoid Revolution

The emergence of models like ChatGPT has shown that AI can process multi-modal inputs. Trained on vast datasets, these models can understand complex relationships between visual features and language, thus enabling robots to navigate real-world tasks more effectively.

For instance, using a model like OWL-ViT, I uploaded an image of an espresso maker and queried it for specific components. The model accurately identified various parts, showcasing its potential for real-world applications.

Furthermore, the multi-modal capabilities of transformer architectures enhance robots' abilities to learn through demonstration. Institutions are increasingly documenting dual-arm manipulation tasks, paving the way for practical applications in robotic assembly.