This challenge provides the opportunity to develop data-driven methods for making predictions to help solve intriguing object rearrangement problems. Object rearrangement is a fundamental problem in many everyday tasks, e.g., tidying up a desk, organizing a shelf, or sorting products. For the same reason, the problem is of key interest in the development of robotic manipulation solutions.
The problem considered in the competition involves a set of objects within a bounded two-dimensional workspace that must be moved from one configuration to another. Only one object can be moved at a time to mimic the setup where the objects are being picked and placed by a robotic arm. Objects are not permitted to collide with each other. The typical objective is to realize a rearrangement with the fewest moves.
Based on prior work, competition organizers have decomposed the rearrangement problem into three classification subtasks or challenges—monotonicity detection, object selection, and buffer selection—each of which can be approached with machine learning tools.
Subtask 1
Classify each rearrangement instance based on monotonicity, that is, whether the rearrangement problem can be solved by moving each object only once.
Subtask 2
Given a specific object o in the environment, predict whether a non-monotone rearrangement problem can be monotone after o is moved away (to an external buffer location).
Subtask 3
Given a specific object o and a buffer location b in the environment, predict whether there is a rearrangement plan where o moves twice and uses b as the intermediate pose while other objects move only once (directly to the goal pose).
Training and test data for the above classification tasks is provided via Kaggle. Please refer to the Kaggle site for more information and to join the competition.
You may also view this PDF document to read a description of the three challenge subtasks, how to access the data, and how to submit solutions.
Kaggle site: https://www.kaggle.com/c/object-rearrangement-competition
This competition is sponsored by DATA-INSPIRE, a TRIPODS Institute based at DIMACS.