With this new technology, robots can really become refined!

With this new technology, robots can really become refined!

According to the news from engadget, the machine vision system released by MIT's latest research can allow the robot to solve the problems it sees by itself.

For machines that perform specific actions repeatedly, the current computer vision technology is sufficient, such as removing an object from the assembly line and placing it in a box. However, for robots to become useful enough not only to load boxes in the warehouse, but also to help in our own homes, the current technology seems to be inadequate.

In response, the Massachusetts Institute of Technology designed the "DON" system.

DON, or "dense object network", is a new form of machine vision developed by the MIT Computer Science and Artificial Intelligence Laboratory (CSAIL). It can generate a "visual road map"-a collection of visual data points arranged in the form of coordinates. This system will stitch these separate coordinate sets into a larger coordinate set, just like your mobile phone can stitch multiple photos into a panoramic image. This allows the system to better and more intuitively understand the shape of the object and how it works in the surrounding environment.

"The roughest and most advanced thing you can get from a computer vision system is to detect targets," Lucas Manuelli, a PhD student, the author of the research paper, told Engadget. "The next step is to mark pixels. All pixels are part of a person or a road. Many self-driving car systems perform both operations."

But if you really want to interact with an object in a certain way, such as grab a shoe or grab a cup in a certain way, he continued, then there is only one bounding box or all corresponding to the cup The number of pixels is not enough. Our system has to get finer details in the object... this kind of information is necessary for more advanced manipulation tasks."

In other words, the DON system allows the robot to correctly locate the handle of the coffee cup when looking at a cup of coffee. When the robot picks up the coffee cup, the DON system allows the robot to realize that the bottom of the cup needs to be kept pointing downward to avoid spilling the contents of the cup.

More importantly, the system will allow the robot to pick out specific objects from a bunch of similar objects.

"Many manipulation methods cannot allow the robot to recognize specific parts of the object in different directions," Manuelli wrote in the study. "For example, the existing algorithms cannot grasp the mug with the handle, especially when the mug faces different directions. Direction."

The system relies on an RGB-D sensor with a combination of RGB depth cameras. Moreover, this system can be self-training-if you want the system to recognize brown boots, you only need to put the robot in a room with brown boots, and then the system will automatically start in a loop to obtain the coordinates used to generate the coordinates. Refer to the photos, and then train yourself based on what you see. The whole process takes less than an hour.

"In factories, robots usually require complex parts to work reliably," Manuelli wrote. "But a system like this that understands the direction of an object can be grasped and adjusted accordingly by just taking a picture."

This technology is still in its infancy, and Manuelli hopes that after continuous improvement, it can be applied to daily household appliances to help people live better.

       AI Rebuild_ai