Researchers from Carnegie Mellon University have been able to teach robots to pick up transparent or reflective objects. This has been a challenge for robots in the past, and the way researchers have addressed the issue is to teach the robots to infer shapes from color images. Picking up transparent and reflective objects has long stymied robots, but the new system promises to alleviate that issue.
One of the coolest features of the new technique developed by the team is that it doesn’t require complicated sensors, lots of training, or human guidance. Instead, it relies primarily on a color camera embedded in the robot arm. Researcher David Held said that death cameras shine infrared light on an object to determine its shape, and that technique works well for opaque objects. The Challenge with transparent and reflective objects is the light either passes straight through or scatters off the surface preventing the depth camera from calculating an accurate shape.
However, a color camera can see transparent and reflective objects as well as opaque ones. The researchers developed a color camera system that’s able to recognize shapes based on color. Using the technique, the researchers were able to train the system to imitate the depth system and implicitly infer shape to grasp objects.
The team used depth camera images of opaque objects paired with color images of the same objects. Once training was complete, the color camera system applied to transparent shiny objects. Based on those images, along with the information a depth camera was able to provide, the system can grasp the challenging objects with a high degree of success.
Held noted that sometimes the arm would miss, but it performed better than any previous system for grasping transparent or reflective objects. The system is still able to pick up opaque objects more efficiently than transparent or reflective items. The system is also able to grasp objects in cluttered piles.