US scientists have given robots a sense of touch through an existing technology. This makes them better at estimating and picking up objects.
Eight years ago, researchers at MIT devised a sensor technology called GelSight. This technology uses physical contact with an object to map its surface. By mounting GelSight sensors on the grippers of robotic arms, the researchers have given robots better sensitivity and dexterity. It tracks objects with point cloud based on vision and touch. The research was presented in two parts.
Smaller objects
In one part, the researchers use data from the GelSight sensor to enable the robot to assess the hardness of the surfaces it touches - a crucial feature when household robots come into contact with everyday objects. In the other part, the sensors are used to allow the robot to manipulate smaller objects than previously possible.
According to the researchers, the GelSight sensor is in some ways a simple solution to a complicated problem. It consists of a block of transparent rubber - GelSight's gel - part of which is coated with metallic paint. When the painted side presses against an object, it moulds itself to the object.
The metallic paint makes the object's surface reflective, making its geometry easier to deduce by computer algorithms. On the sensor opposite the coated side of the rubber block are three coloured lights and a camera. "The system has coloured lights aimed at different angles and the reflective material. By looking at the colours, the computer can determine the 3D shape of the object."
Algorithms
For an autonomous robot, 'feeling' the softness or hardness of an object is essential to determine where and how it should pick it up but also how the object behaves when it is moved, stacked or placed on different surfaces. It normally does this using a computer vision system. Such systems provide very reliable information about an object's location - until the robot picks it up. Especially if the object is small, much of it will be enclosed by the robot's gripper, making determining a location more difficult. This makes the robot's estimation unreliable, just when the it should know the object's location exactly. This also posed a problem when the MIT team had the robot pick up a drill and turn it on; that took two to three minutes. "So it would be much better if we had a real-time accurate estimate of where the drill was and where our hands were relative to the drill."
Consequently, the team has developed control algorithms that use a computer vision system to guide the robot's gripper as it approaches a tool and then relays the location determination to a GelSight sensor once the robot has the tool in its hand. Generally, the challenge with an approach like this is that the data produced by the vision system has to be connected to the data produced by a touch sensor. But GelSight has its own camera, so its output is easier to integrate with visual data than the data from other touch sensors.
Screwdriver
In one part of the experiment, a robot equipped with a GelSight gripper had to pick up a small screwdriver, remove it from a holster and put it back. The data from the GelSight sensor did not describe the whole screwdriver, only a small part of it. But the researchers found that as long as the vision system determined the original position of the screwdriver within a margin of a few centimetres, its algorithms could deduce which part of the screwdriver was touching the GelSensor. In this way, the position of the screwdriver in the robot's hand could be determined.
"I think the GelSight technology, like other tactile sensors, is going to have a big impact on robotics," argues a professor from the University of California. "In humans, sense of touch is one of the key factors for our fantastic manual dexterity. Finding the light switch in the dark by touch, taking something out of your pocket or other things we do without thinking about it all depend on our sense of touch. Existing robots lack this type of dexterity and are limited in their ability to respond to surface features when dealing with objects. But now software is catching up with sensors and more is possible."
By: Kelly Bakker
Source: MIT
