In the world of industrial robotics, it’s easy to fall into a way of thinking in which the hardware and software you need to solve your automation challenges lives only in the pages of a catalog from ABB, FANUC or another large corporation. The truth is, there are plenty of “little guy” companies out there building innovative new products with the motors, mechanisms and code that makes industrial robots tick.
CapSen Robotics is a small company of five engineers based in Pittsburgh, PA. The company was formed in 2014 by CEO Jared Glover after he began looking to commercialize his PhD research at MIT for detecting known 3D objects in cluttered 3D images. I spoke to Glover about that research, his new company and its products.
“We settled on bin picking because it’s a good fit for this technology,” he said. “Most bin picking problems are with known rigid objects. In most typical cases, you have—or can acquire—3D models of the objects. The challenge is that they’re jumbled in the bin. A bin-picking robot can pick them out one at a time.”
The team started by tuning up Glover’s research code and progressed to developing their core offerings: bin picking software, a plug and play system including software, cameras and computer; and a turntable 3D scanner.
Challenges of Developing Custom Bin Picking Software
“The biggest challenge is that it’s a very big search problem,” Glover explained. “If you’re searching for something in an image flat on a conveyor belt, you are searching in dimensions x and y, plus orientation. But in a bin, objects can be at any x, y or z coordinate location and any orientation, so there are six dimensions to search over. It’s a lot harder to do that search efficiently.”
The second challenge CapSen’s system faced was cycle time. Processing object detection is intensive, and this is one of the key areas of innovation in the bin picking game. To gain an edge, CapSen is using a technology called Nvidia CUDA GPU processing.
“One of the things we spent a lot of time building was a generic programming library we call multimatrix,” Glover said. “This performs computations on sets of matrices.”
CapSen’s underlying software infrastructure, this multimatrix library, is similar to MATLAB except it’s optimized for the kinds of computations in computer vision and robotics.
Supported Robots, Cameras and End Effectors
The company has installed its system on ABB and Mitsubishi robots. However, like most bin picking products on the market, CapSen aims to be completely device agnostic. According to Glover, the process of adding support for a new robot takes about three weeks. The company works with customers to add support for its devices as needed.
The type of 3D camera used will impact the type of object you can recognize. For example, if you use Kinect, it won’t detect parts the size of your fingernail. CapSen can recommend the appropriate camera for each customer’s application.
The most common end effector for bin picking, and indeed most pick and place, is suction cup-based. However, finger grippers are sometimes required when dealing with certain parts. For example, picking porous or irregular parts is difficult or impossible with suction.
Making Bin Picking Robotics Accessible for Non-Technical Personnel
In order to be useful for small and medium enterprises, especially those with high-mix, low-volume production runs, Glover knows his system must be usable and configurable by non-technical personnel. Here’s how his system achieves this goal:
To set up a bin picking cell, the software needs a 3D model of each object to be picked. However, not all parts have a digital model, and 3D modeling is a highly skilled and time-consuming job. A growing workaround to this problem is the turntable 3D scanner, which can generate a 3D model as easily as microwaving a burrito. CapSen offers one of these scanners as part of this strategy.
“There are some competitors who sell the camera and software to find the objects,” Glover said. “But that still places the burden on the integrator to move the robot and find the object. We’re trying to make that process as easy as possible.”
Next-Level Applications for Bin Picking Technology
CapSen continues working on functionality, including support for multiple objects in the same bin, support for picking from a bin in an unfixed position and faster cycle time. In addition, Glover has more ideas for software applications.
One idea is mobile material handling, which is essentially bin picking on a mobile base.
“Rather than picking little parts, it’s boxes or barrels,” Glover said. “This is an exciting area that we look forward to getting into.”
Mobile material handling could have big implications in the shipping and warehousing industries, where thousands of human pickers currently scurry around on pallet jacks and forklifts like ants.
Another potential application is improved machine vision for assembly robots. Robots that use multiple tools for a task require the tools to be arranged in highly accurate fixturing because the robot accesses the tool using coordinate programming. However, if assembly robots could use vision-based picking software, they could manipulate workpieces and tools with more versatility and flexibility.
Innovation from small companies, even those comprised of just a few talented engineers, can rival that of the large multinational technology companies. In this case, it all started with a PhD research project.
If you have an idea for a project, why not start it up on Projects.engineering.com?