At the start of the COVID-19 pandemic, auto companies like Ford quickly shifted their production from cars to masks and ventilators.
To make this transition possible, these companies relied on people working on the assembly line. It would be too difficult for a robot to make this transition because robots are tied to their normal tasks.
In theory, the robot could lift almost anything if its grips could be swapped for each task. To keep costs down, these grippers can be passive, meaning the grippers grip objects without changing shape, similar to how grippers on a forklift work.
A team at the University of Washington has created a new tool that can design a passive gripper for 3D printing and calculate the best path to pick up an object. The team tested this system on a set of 22 objects – including a 3D-printed rabbit, a wedge-shaped door stop, a tennis ball and a drill. Designed captures and trajectories were successful for 20 objects. Two of them were wedge-shaped and a pyramid with a curved keyhole. Both forms are complex for different types of grips.
The team will present these results on August 11 at SIGGRAPH 2022.
“We still make most of our products on assembly lines, which are really great, but also very tough. The pandemic has shown us that we need a way to easily repurpose these production lines,” said senior author Adriana Schultz, an associate professor at UU. in the Paul G. Allen School of Computer Science and Engineering. “Our idea is to create a custom tool for these production lines. This gives us a very simple robot that can perform a single task with a specific grip. And then when I change the task, I just replace the grip.’
Passive grabs cannot adapt to the object they grab, so traditionally objects are designed to match a specific grab.
“The most successful passive grip in the world is the forklift tongs. But the trade-off is that the forklift tongs only work well with certain shapes like pallets, which means that whatever you want to grab must be on the pallet, ” said co-author Jeffrey Lipton, an assistant professor in the UW Department of Mechanical Engineering. “Here we say, ‘OK, we don’t want to predetermine the geometry of the passive capture.’ Instead, we want to take the geometry of any object and design a grip.”
For any given object, there are many options for what its capture might look like. In addition, the grip shape is related to the way in which the robot arm picks up the object. If not designed properly, the gripper can bump into the object on the way to pick it up. To solve this problem, the researchers had several key findings.
“The points at which the gripper makes contact with the object are important for maintaining the stability of the object in the grip. We call this set of points the ‘grip configuration,'” said lead author Mylin Codnongbua, who completed this research as a UW undergraduate in Allen School. “Furthermore, the gripper must contact the object at these specified points, and the gripper must be the only solid object connecting the contact points to the robot arm. We can look for an insertion path that satisfies these requirements.”
When designing a new grip and trajectory, the team begins by providing the computer with a 3D model of the object and its orientation in space—for example, how it will be presented on the assembly line.
“First, our algorithm creates possible capture configurations and ranks them based on stability and some other metrics,” Codnongbois said. “Then it takes the best fit and performs co-optimization to determine if the insertion path is feasible. If it can’t find it, it moves to the next capture configuration in the list and tries to co-optimize again.”
Once the computer has found a good match, it issues two sets of instructions: one for the 3D printer to create the grip, and another with a path for the robot arm once the grip is printed and attached.
The team chose a variety of objects to test the power of the method, including some from a dataset of objects that are the standard for testing a robot’s ability to perform manipulative tasks.
“We also designed objects that would be difficult for traditional grasping robots, such as objects with very small angles or objects with internal grasping – where you have to pick them up by inserting a wrench,” said co-author Ian Good , a UW doctoral student in the Department of Mechanical Engineering.
The researchers conducted 10 test pickups with 22 shapes. For 16 forms, all 10 pickups were successful. While most figures had at least one successful pickup, two did not. These failures were the result of problems with the 3D models of the objects that were transferred to the computer. For one – the bowl – the model described the sides of the bowl as thinner than they were. For another – a cup-like object with an egg-shaped handle – the model did not have the correct orientation.
The algorithm developed identical grasping strategies for similarly shaped objects, even without human intervention. The researchers hope that this means they will be able to create passive grippers that can grab a class of objects, instead of having a unique gripper for each object.
One limitation of this method is that passive captures cannot be designed to capture all objects. While it is easier to pick up objects of varying widths or with protruding edges, objects with equally smooth surfaces, such as a water bottle or box, are difficult to grasp without any moving parts.
However, the researchers were encouraged that the algorithm worked so well, especially with some of the more complex shapes, such as a column with a keyhole at the top.
“The way our algorithm came up with that was to accelerate rapidly to where it was approaching the object. It looked like it was going to crash into the object, and I thought, ‘Oh, no.’ What if we didn’t calibrate?'” Hood said. “And then, of course, he gets incredibly close, and then captures it perfectly. It was this extraordinary moment, an extraordinary roller coaster of emotions.”
Yu Lu, who completed this research as a graduate student at the Allen School, is also a co-author of this work. This research was funded by the National Science Foundation and a grant from the Murdock Charitable Trust. The team has also applied for a patent: 63/339,284.
