The sensor is an adaptation of a technology called GelSight, which was developed by the lab of MIT's Professor Edward Adelson, and first described in 2009. The new sensor isn’t as sensitive as the original GelSight sensor, which could resolve details on the micrometre scale, but it’s smaller — small enough to fit on a robot’s gripper — and its processing algorithm is faster, so it can give the robot feedback in real time.
Industrial robots are capable of remarkable precision when the objects they’re manipulating are perfectly positioned in advance. But according to Robert Platt, an assistant professor of computer science at Northeastern University and the research team’s robotics expert, for a robot taking its bearings as it goes, this type of fine-grained manipulation is unprecedented.
“People have been trying to do this for a long time,” Platt says, “and they haven’t succeeded because the sensors they’re using aren’t accurate enough and don’t have enough information to localise the pose of the object that they’re holding.”
Whereas most tactile sensors use mechanical measurements to gauge mechanical forces, GelSight uses optics and computer-vision algorithms. A GelSight sensor — both the original and the new, robot-mounted version — consists of a slab of transparent, synthetic rubber coated on one side with a metallic paint. The rubber conforms to any object it’s pressed against, and the metallic paint evens out the light-reflective properties of diverse materials, making it much easier to make precise optical measurements.
In the new device, the gel is mounted in a cubic plastic housing, with just the paint-covered face exposed. The four walls of the cube adjacent to the sensor face are translucent, and each conducts a different colour of light — red, green, blue, or white — emitted by light-emitting diodes at the opposite end of the cube. When the gel is deformed, light bounces off the metallic paint and is captured by a camera mounted on the same cube face as the diodes.
From the different intensities of the different-coloured light, the algorithms developed by Adelson’s team can infer the three-dimensional structure of ridges or depressions of the surface against which the sensor is pressed.
Although there are several ways of measuring human tactile acuity, one is to determine how far apart two small bumps need to be before a subject can distinguish them just by touching; the answer is usually about a millimetre. By that measure, even the lower-resolution, robot-mounted version of the GelSight sensor is about 100 times more sensitive than a human finger.
In Platt’s experiments, a Baxter robot from MIT spinout Rethink Robotics was equipped with a two-pincer gripper, one of whose pincers had a GelSight sensor on its tip. Using conventional computer-vision algorithms, the robot identified the dangling USB plug and attempted to grasp it.
It then determined the position of the USB plug relative to its gripper from an embossed USB symbol. Although there was a 3-millimetre variation in where the robot grasped the plug, it was able to measure its position accurately enough to insert it into a USB port that tolerated only about a millimeter’s error.
“Having a fast optical sensor to do this kind of touch sensing is a novel idea,” says Professor Daniel Lee from the University of Pennsylvania and director of the GRASP robotics lab. “I think the way that they’re doing it with such low-cost components — using just basically coloured LEDs and a standard camera — is quite interesting. This is a promising prototype. It could be developed into practical device.”