Servo - July 2010

GEERHEAD

Here, the Aqua robot and a diver are working together in a pool. The robot obeys image-based commands delivered by hand in front of its camera. This control technology was developed because the robot cannot receive radio-based commands underwater.

These image-based communications are called RoboChat. All underwater communications must be acoustic or visual. Acoustics are typically too slow. Some of the commands the diver gives the robot are stop, go, follow the diver, surface, etc. The technology also enables the diver to give the robot numerical inputs. The robot can be programmed while under water using these images.

follow a diver or the contours of a reef,” explains Dudek. “It learns patterns associated with a particular diver it wants to follow, and looks for those patterns and follows them.” These patterns include combinations of colors, motions, and textures.

This is called target recognition. The technology uses a filter that analyzes each picture the robot takes. If it is tracking a face, it will have been trained on what two eyes look like and it will search for those. The technology uses hundreds of filters. This is called ensemble tracking because it uses an ensemble of filters.

In the first stage of training the robot’s tracking system, researchers show the robot an image of something they want it to follow. In the second stage, some of the robot’s systems train the other systems. The robot’s interworkings constantly adapt to what one of its systems knows is a correctly identified target.

In the research at York University, the robot applies trinocular sensors which use two cameras together to see how far apart objects are under water. One camera looks ahead following a diver, while another looks at the terrain below, mapping it out.

Thanks to the computers and cameras, the robot takes better images including videos (watch the lobster video in the resources). When the robot surfaces, researchers may want to know right away if it saw anything of significance. The robot may have taken three hours of video, but with its new technology it can identify the 25 most interesting images very quickly.

The problem with selecting the things that are interesting from all the images is that the technology has to model things that people would find interesting. So, the robot needs to be able to identify those. In the model, some things are interesting because they are new, and

12 SERVO 07.2010

some are interesting because they are unlike any previous images.

The Aqua robot works in tandem with stationary sensors in the water. There is a separate project dedicated to deploying these sensors to collect data. An Aqua robot can swim to a sensor, pick up the data, and retrieve it.

The Aqua robot also has legs — a new set that are fully amphibious, so it can walk and swim, and do both very well. Photos are currently unavailable as this technology is still a bit under wraps. Having the Aqua robot swim and then walk directly onto shore when done opens up a lot of possible uses. These include military applications where the bot can stay on the beach and then go into the water to de-mine the surf zone. The robot could also walk into the water, swim and collect data, crawl back out of the water, and walk into a designated area and uplink the data. Or, it can follow a diver into the water.