Realtime Concurrent Planning and Plan Execution in Stochastic Domains

Luis Pineda and Shlomo Zilberstein. Realtime Concurrent Planning and Plan Execution in Stochastic Domains. Technical Report 2014-21, School of Computer Science, University of Massachusetts Amherst, 2014.

Abstract

In realtime planning domains, such as service robot control, an agent receives a task and must minimize the combined cost of planning and plan execution necessary to complete the task. To reduce the total cost, we examine the feasibility of performing planning continuously, while parts of the intermediate plan are being executed. The main challenges are to guarantee the completeness of the approach and make sure that planning does concentrate on regions of the state space that are most crucial given the state of execution. Surprisingly, simple modifications of existing stochastic planners yield an efficient approach for concurrent planning and plan execution. We formalize this approach and analyze its characteristics. Experimental results show that such a continuous planning paradigm offers significant benefits, most notably a significant cost reduction relative to existing realtime planning and execution strategies.

Bibtex entry:

@techreport{PZtr1421,
  author	= {Luis Pineda and Shlomo Zilberstein},
  title         = {Realtime Concurrent Planning and Plan Execution in
                   Stochastic Domains},
  year          = {2014},
  number        = {2014-21},
  institution   = {School of Computer Science, University of Massachussetts Amherst},
  url		= {http://rbr.cs.umass.edu/shlomo/papers/PZtr1421.html}
}

shlomo@cs.umass.edu