Scientists from the Massachusetts Institute of Technology (CSAIL) Computer Science and Artificial Intelligence Laboratory have developed a new type of multifunctional robotic machine. They say this machine can learn to pick up and drop off all kinds of things.
Keypoint Affordance Manipulation (KPAM) is control software, designed to offer robots much greater flexibility.
This project aims to perform the manipulation of a reflective robot in its movements for a variety of objects.
But the CSAIL researchers say that a typical manipulator relies primarily on estimating the position and orientation of objects and using geometry-based algorithms to capture them. Although there is restriction, especially when it comes to objects of very different shapes and it is about repositioning them in a subtle way.
The method is based on a set of key points on the object, which it interprets as coordinates. They took the cup as an example, the system only needs three coordinates to grab, the key points are located on the side, bottom and center of the handle, which are enough to complete the task.
A lot of hardware experiments show that this method can reliably perform tasks for objects that have never been seen in the category, such as placing shoes and cups when the category-level goal setting changes a lot.
The experiment included the Kuka IIWA LBR robot clamped with the Schunk WSG 50 clamp (and the Primesense depth detection sensor), the robotic model delicately dissected the shoe in the shoe rack, placed upright on the shelf. And put a mug on the shelf through the handle.
By powering the system in six degrees of freedom, the scientists were able to make the Kuka IIWA LBR robotic arm running KPAM (collect more than 20 different shoes, including everything from athletic shoes to boots. Although it had some trouble lifting a pair In high heels, adding a few pairs to his neural network training data allowed him to quickly complete the task.
In the process of the cup holder, this includes 40 cups with different shapes, sizes and visible appearances. The robot controls all the cups to grab them in the vertical position, but it is better to grab 19 cups when in the horizontal position. It is because the movement of the gripper is restricted. What is impressive is that he placed the mug on a shelf 5cm from the target position in all attempts.
of 30 cups 5 were very small with handles of 2 cm wide, the robot managed 100% of the time to hang larger cups on the shelf, as for the small ones it achieved 50% effectiveness, The team of researchers believe that failures are due to wrong key points
Finally, this robotic tool together with its software for manipulating objects with key points in 3D is based on an optimized robot action plan and performing functions supported by capturing with a more effective vision.