Formal synthesis of optimal information-gathering policies

This paper considers the problem of informative path planning under temporal logic (TL) correctness constraints. For example, a robot deployed to a building after a natural disaster must explore the area and report possible locations of survivors with minimum uncertainty. The robot must satisfy the correctness requirement “Always avoid obstacles. Report data to rescuers at a data upload hub before exiting the scene. If a location with suspected fire is visited, investigate and report immediately.” In this work, we map the constrained informative path planning problem to a stochastic optimal control problem over a Markov decision process that integrates a robot’s motion under a given TL specification with its sensing model. We develop an optimal dynamic programming algorithm and a receding horizon approximation with significantly reduced computational cost. Both solutions are guaranteed to satisfy the given specification and are evaluated using simulations and experiments with ground robots. Our results show that the receding horizon solution approximates the optimal method closely, which indicates usefulness for persistent information gathering tasks.