Executive Summary : | Recent advances in computing and communication technologies have made engineering systems more complex and expected to perform complex tasks. For example, a drone air delivery can be modeled using Spatio-temporal logic formulas like Linear Temporal Logic (LTL)¹ or Signal Temporal Logic (STL). Researchers have been using symbolic control techniques to design controllers enforcing these complex tasks. However, there is limited literature available that utilizes learning-based approaches for enforcing complex tasks represented using LTL or STL. There is an urgent need for developing various learning-based controller design approaches for partially/fully unknown dynamical systems enforcing complex Spatio-temporal logic tasks. The project aims to provide solutions to this problem by proposing reinforcement learning-based approaches to learn control policies that enforce Spatio-temporal tasks. The solutions focus on two main categories: end-to-end learning of controllers and learning approximate unknown dynamics then designing controllers enforcing these tasks. These learning-based techniques are computationally heavy and require high-performance computing facilities. To achieve this, the project aims to utilize in-house supercomputing clusters available at IISc. The project also plans to demonstrate the proposed results on control of autonomous ground/aerial vehicles in virtually created urban-like environments. They have a motion capture system with multiple cameras and communication units to collect high-quality initial training data and a projection system that projects a virtual environment for robots, imposing complex Spatio-temporal logic specifications. |