TUBITAK 1001 Programme Project Number: 122E249
A data-driven approach using temporal logics for repairing robot controllers
Budget: 80.000 dolar
Project Summary : Robots are used in many different areas such as housework, production lines and defense industry. In general, a robot that is designed to perform a variety of tasks can achieve a specific one via the controller designed for the task. Controller design is a very challenging problem that requires taking into account the constraints and the properties of the considered robotic platform, the given task, the related optimization criteria, and the environment that the robot will operate in. In this project, the goal is to develop a novel data-driven approach using temporal logics for robot controller design and to apply it on two robot platforms. The main steps of this approach are (1) generating a labeled dataset by simulating the robot with a set-valued controller, (2) generating temporal logic formulas describing positive and negative events from the dataset, and finally (3) repairing the controller such that the trajectories generated with the repaired controller violates (or satisfies according to the label) the formula. Furthermore, these steps can be repeated to improve the resulting controller. Traditionally, temporal logics are used for verifying software and hardware systems. Due to their expressivity and existence of algorithms that can be adapted for different tasks, they have gained popularity in other areas including robot control and analysis of dynamical systems. In this project, temporal logic formulas will be used to define temporal patterns that lead to positive and negative events. In order to implement the proposed approach, controllable formula templates that can be used for modifying the controller to satisfy/violate the formula and the associated repair processes will be defined. In addition, a machine learning based method will be developed to solve the formula synthesis problem. In addition to generating controllable formulas from the labeled datasets, this method will be applied to the datasets published together with the recent works on formula synthesis in order to compare the results in terms of the classifier success and computation time. The first step of the proposed approach is simulating the robot with a set-valued controller. This controller can be developed with the best effort principle, it can be an exploratory controller applying only control constraints, or it can be a controller developed via machine learning. In this project, two controller learning methods based on reinforcement learning and deep learning from the literature will be implemented. The learned controllers that do not fully comply with the constraints or do not fulfill the task will be used to produce a labeled dataset in the first step and the repair approach will be applied to these controllers. On the other hand, controllers that offer a solution will be used to evaluate the performance of the proposed approach. In addition, in order to understand the working principle of the learned controllers, the formula synthesis method will be applied to the datasets generated with these controllers. Consequently, as an additional benefit of the proposed approach, the working principles of the learned controllers will be expressed in an understandable way with temporal logics. The proposed approach can be applied to various platforms since it is data-driven and does not directly use the system dynamics. Furthermore, it is expected to provide a robust controller against uncertainties that is inevitable in physical systems. In order to confirm these and solve some problems encountered in the control of unmanned surface robots, this approach will be applied to develop controllers from motion planning tasks for an unmanned land vehicle and an unmanned surface robot. Throughout the project, the proposed methods will be applied in simulation environments and on the real robots for both of the platforms.
TUBITAK 1001 Programme Project Number: 118E195
Feedback Motion Planning of an Unmanned Surface Vehicle via Random Sequential Composition
Budget : 70.000 Dolar
Project Summary : In this project proposal, our goal is to develop an original unmanned surface vehicle (USV), and a novel methodology that solves the path planning and navigation problems for this system (and other USV systems) in a robust and computationally feasible way. Our approach in solving the motion planning problem is based on synthesizing two different existing methodologies. One of these methods is sampling based path planning, and the other one is the sequential composition of dynamic behaviors. Our aim for adopting such a synthesis route is to create a hybrid framework that exploits the strengths of two different approaches and removes the possible weak points of each method. In this project proposal, first we plan to develop the algorithm and test the results in a simulation environment. In parallel to these simulation studies, we aim to develop a new USV that can be used effectively in robust motion planning and navigation studies. After the simulation studies and full development of USV, we are planning to implement and test the methods on the developed USV system using field experiments.