A new algorithm enables robots to pick up objects much faster. The algorithm helps robots decide the strategy they use for this in just seconds.
The algorithm was developed by researchers at the Massachussets Institute of Technology (MIT), who present the results of their development in The International Journal of Robotics Research. Picking up an object is a breeze for humans. For robots, however, the same action is a very complex movement. Thus, all kinds of factors affect the robot's ability to pick up an object, such as the geometry of the object and the friction provided by the material.
Plans
To successfully pick up an object, then, a robot must plan its action to a significant extent. One way to pick up such an object is to push the object against a solid surface. Researchers at MIT have found a way to significantly speed up the planning process required for this. Whereas traditional algorithms can easily take 10 minutes to plan such an action, MIT's algorithm can do it in less than a second.
Part of the reason why traditional algorithms take a relatively long time to plan actions is that a variety of natural laws have to be taken into account in a calculation. These include Newton's laws of motion, but also Coulomb's Law to determine friction between objects. "It is a long-winded computational process to integrate these laws, consider all the possible movements the robot can make and choose the most useful one from these," Rodriguez said.
Using environment
By being able to plan faster, robots in industrial environments, among others, can more quickly determine how to use their environment to manipulate objects. For example, they can push objects against a surface or slide them along an edge to manipulate the object's orientation. Alberto Rodriguez, assistant professor of mechanical engineering at MIT, calls this useful for picking up and sorting objects, as well as for accurate use of tools, among other things. "This is a way to increase the skills of even simple robotic grippers, regarding every robots environment at its disposal," Rodriguez explains.
The team's method uses 'motion cones', which are basically cone-shaped maps for friction. The inside of the cone represents all the pushing motions that can be applied to an object in a specific location within fundamental laws of nature to hold it. The space outside the cone represents all the pushing motions that result in the robot losing its grip on the object. The algorithm does not have to perform the calculations required to determine this thanks to the cone for every movement.
'Complicated process'
"Seemingly simple variation such as how hard the robot grabs an object can fundamentally change the way an object moves if pushed. This is part of the physical reasoning that the algorithm takes care of," explained Rachel Holladay, an electrical engineering and computer science student at MIT. "It's a complicated process, but much faster than the traditional method - fast enough to plan a full sequence of pushing movements in half a second."
The algorithm was tested in a test setup where a simple robot arm picks up the T-shaped block, pushes against a vertical surface to rotate the block and places it next to the letters M and I. The robot is presented with the block in a different orientation each time during the test. The algorithm was able to instantly create a map of all the possible forces the robot can apply to the block and the impact these movements have on the block's orientation.
"We made thousands of pushing moves to verify that our model correctly predicts what happens in the real world," Holladay said. "If we make a movement that is inside the cone, the grasped object should remain under control. If it is outside this, the object should escape from the grip."
Removing constraints
"Many factory robots that use tools have a specially designed hand. Instead of having the ability to pick up a screwdriver and use it in different ways, they simply turn the hand into a screwdriver," says Holladay. "You can imagine that this requires less detailed planning but creates more constraints. We want a robot capable of picking up many different things."
Author: Wouter Hoeffnagel
Source: MIT
