MACS project aims to revolutionize robotic perception
April 28, 2008 We’re getting better at making robots that look like us, and move creepily like us, (or just plain creepily), but attempting to program our perception and reasoning is proving to be an order of magnitude harder. The Multi-sensory Autonomous Cognitive Systems (MACS) project is an attempt to imprint robots with the ability to understand the affordances of objects in their environment - or the physical qualities of an object, and the tasks it can be used for. It is hoped that this new approach to computer perception will allow robots to perceive more interaction possibilities, giving them the cognitive tools they need to successfully improvise and interact with their environment in a more human manner.
A lot of roboticists have applied the same design concept to robot perception as they have to robot appearance – if it looks like us, and moves like us, it’s close to being us. The Kismet robot, for example, has high quality cameras in place of eyes, and pattern recognition software in place of comprehension, but the result is just a sophisticated aping of human interaction. It seems increasingly clear that the path to Artificial Intelligence doesn’t lie in mimicking the magnificence of the human biology, but rather the subtlety of the human psychology.
While Computer Vision may provide rich visual data and appearance-based algorithms can allow robots to identify objects, the approach appears flawed in its restrictiveness. The fact that a robot might identify a glass as a vessel for water, but not a mug, illustrates the fundamental futility of the system – however many objects are added to its database, the robot will still be imitating our comprehension, not sharing it. Alternative means of identification - like simply using a pointer - are in development, but again these methods have limits.
The MACS “affordances-based” strategy expands robot perception beyond simple recognition – instead of being appearance-based, it is feature-based. A traditional robot may be able to search its environment for a particular object it recognizes as a chair, but a MACS robot will search its environment for a particular object that could allow it to be seated – whether it be a chair, stool, couch or ledge. This comprehension enables a human-like approach to tasks – if sent to get water, a MACS robot would identify glasses, cups, and mugs as appropriate vessels without relying on restrictive definitional parameters. A goal-setting robot would be able to use its affordance-based perception to efficiently complete its task with materials at hand, improvising like a human.
The EU-funded MACS research began in 2004, with five goals: to create the necessary software architecture for affordance-based robot control; to use this software to direct a robot to complete a goal-directed task; to establish methods for reasoning about affordances; to allow robots to learn new affordances through experimentation; and to demonstrate the entire system on a robotic platform called the Kurt3D.
Kurt3D can use the MACS affordance-based perception to identify what can be grasped, what can be traversed, and where there was free space. It demonstrated this by finding an object, picking it up, placing it on a pressure-activated switch, and moving through a door.
“This is the very early stages of this approach,” warns Dr Erich Rome, director of the MACS project. “So we are a long way from commercialization. There are others working on it. But what is unique about the MACS project is that we introduced direct support for the affordances concept in our architecture.”
For upwards of fifty years, scientists have had difficulty even defining what Artificial Intelligence is, let alone how we may reach it. As milestones have been reached in processing ability, the goalposts have shifted away from logic and calculation, and more toward perception - any old Mac can beat a human at chess nowadays, but few would say it was actually "smarter". Increasingly, emphasis has been placed on how well a computer can interact with its environment, how creative it can be in achieving goals, its ability to experiment and learn, and its ability to improvise. By changing the way machines view their environment, the MACS project may have laid the foundations for a truly intelligent machine.
Further information including explanatory videos can be found at the MACS site.