Mobile Autonomous Robot Software

This is the home page for the Defense Advanced Research Project Agency (DARPA) Mobile Autonomous Robot Software (MARS) project at the Georgia Institute of Technology (GT).

The GT MARS project involves multi-level learning in hybrid deliberative/reactive mobile robot architectural software systems. We are developing revolutionary learning methods in the context of proven DARPA-developed software to accomplish complex adaptive autonomous mobile robot systems in dynamic, hostile environments. We are integrating the following learning modules (both deliberative and reactive) into our existing MissionLab software:

  • Probabilistic situational recognition and indexing into behavior sets for opportunistic planning and reaction.
  • The use of case-based reasoning (CBR) methods for situation-dependent behavioral gain and assemblage switching at the reactive level of execution.
  • The use of CBR for reasoning in Finite State Automata (FSA) plan generation through the use of "wizards" to guide high level deliberative planning.
  • Specialized reinforcement learning methods ("learning momentum") to adjust behavior gains at run-time.
  • The integration of Q-learning methods for behavioral assemblage selection at a level above gain adjustment.