Task level strategies for robots

Abstract

This thesis addresses the problem of constructing strategies to solve robot tasks. We use a planner along with a set of local task-level feedback controllers to create strategies that are robust and have globally convergent behavior. The planner and the local controllers can be interleaved during execution. The information gathered by the local controllers enables the creation of more robust paths, and the path ensures that the local controllers make progress globally. These local task-level feedback controllers are automatically created and updated from simulation models or from empirical trials. They handle the uncertainty and possibly time-varying dynamics that may be present during task execution. We present implementations of this approach in the planar pushing domain and in the non-holonomic planning and control domain. Thesis Supervisor: Tom'as Lozano-P'erez Title: Professor 3 4 Acknowledgements This thesis has taken far longer than it should have, and it might have taken e...