Lifelong Robotic Object Discovery
Robots operating in dynamic environments need to gather information about their surroundings; the better the perception algorithms are, the better and more complex the tasks can be. In the case of personal robotics, learning information about novel objects in the environment is necessary for long-term operation. In this work, our goal is to analyze the input data stream of a robotic system while the robot operates, for as long as the robot operates, and produce 3D models of the novel objects that the robot sees while in operation. We refer to this problem as Lifelong Robotic Object Discovery. In Lifelong Robotic Object Discovery, our algorithms must concentrate on the quality of objects as well as the scalability of the system, given that we are to process potentially days, months or even years of sensory information.
In this work, we discover novel objects from a robot's data stream and generate 3D models for the objects in real time, with no previous knowledge of any object or the environment. State-of-the-art techniques in Object Discovery use graph-based approaches and clustering techniques to describe what objects are. An object is often described as a group of regions from multiple images with strong pairwise similarity between them. The key contribution of our solution is the generation of Constrained Similarity Graphs (CSG) for Object Discovery. Constrained Similarity Graphs encode constraints as multiple sources of similarity in the graph generation procedure. The resulting CSG is orders of magnitude sparser than pairwise similarity graphs and often contains many disjoint components, enabling parallelization. Our solution processes a dataset of over 4000 images in around two minutes, allowing for real time operation.
|Examples of discovered objects with our algorithm.|
This material is based upon work partially supported by the National Science Foundation under Grant No. EEC-0540865.