When humans and robots share the same physical spaces and interact in close proximity to one another, special considerations arise for maintaining the safety and efficiency of the interaction. During such close-proximity interaction, there is increased potential for collisions and motion conflicts, making it imperative that robots are able to anticipate their human partners and adapt accordingly to minimize these negative impacts.
The general drawback of current human prediction approaches, however, is that they are often designed for specific tasks or motions and therefore do not generalize well. This makes it difficult to implement human prediction in practice, as it might not be obvious which prediction approach would perform best in a given task.
The aim of this work, therefore, is to build upon prior work in the field of human intent prediction in order to develop a data-driven approach that can learn from given task data and automatically synthesize a prediction solution that is generalizable and robust.
Namely, given a variety of data encoding how the person moves in the shared environment and how he or she performs the tasks, we have developed a method that automatically selects a favorable combination of prediction methods to accurately predict occupancy at various future timeframes. We then use these predictions as an input to robot motion planning algorithms in order to allow robots to reason about where co-located people will be in the future and select paths and motions that allow for safe and efficient sharing of the space.