In the video below, the lime-green Qbo transits to the "Object Recognition" state from its internal state machine, to learn to recognize its mirror image, as if it were a regular object. Technically, how does Qbo do it?
According to it's YouTube Description:
The "Object Recognition" state is implemented to execute simultaneously different ROS nodes: one responsible for the head and base movement; another to select an object from the image using the stereoscopic vision; another to recognize objects or learn new ones. The object recognition algorithm uses "SURF" descriptors and the "Bags of Words" approach (through the OpenCV library), and stores images in Qbo's internal storage system.
Qbo has several stored answers and behaviors in an internal knowledge base, that we upgrade as the projects evolves, to make questions or orders to Qbo such as "What it this? or "Do this". Qbo interprets the object "Myself" as a an ordinary object, for which it has special answers in its internal knowledge base such as "Woah. I'm learning myself" or "Oh. This is me. Nice". Qbo selects its reflection in the mirror in the image that he sees using the stereoscopic vision, and one of our engineers interacts (speaks) to him so that Qbo can learn to recognize himself as another object. For direct interaction, Qbo uses the open-source software Julius for speech recognition (in the video, you see how Qbo receives the order to turn around, and he responds to it by moving its base 90 degrees), and Festival for voice synthesis.
1 comment:
It's cute, yet kinda creepy. Data's great great great grandfather. It's so technologically advanced and probably costs millions to make. So what's the best idea ever?!! Place robo on a much wider table.
Post a Comment