One of the big dreams of AI has been to build an artificially “intelligent” robot. A robot capable of interacting with people and performing many different tasks. We have seen remarkable progress on many of the component technologies necessary to build AI robots. All these tremendous advances beg the obvious question: Why don’t we have a single example of a truly multi-purpose robot that would, even marginally, deserve to be called artificially intelligent?
I believe the key missing component is representation. While we have succeeded in building special purpose representations for specialized robot applications, we understand very little about what it takes to build a lifelong learning robot that can accumulate diverse knowledge over long periods of time. And that can use such knowledge effectively when deciding what to do. It is time to bring knowledge representation and reasoning back into robotics. But not of the old kind, where our only language to represent knowledge was binary statements of (nearly) universal truth, deprived of any meaningful grounding in the physical world.
We need more powerful means of representing knowledge. Robotics knowledge must be grounded in the physical world, hence knowledge acquisition equals learning. Because data-driven learning is prone to error, reasoning with such knowledge must obey the uncertainties that exist in the learned knowledge bases. Our representation languages must be expressive enough to represent the complex connections between objects in the world, places, actions, people, time, and causation, and the uncertainty among them. In short, we need to reinvent the decades-old field knowledge representation and reasoning if we are to succeed in robotics.