TEAMING.AI Human-AI Teaming Platform for Maintaining and Evolving AI Systems in Manufacturing
Partners and Supporting Agencies: EC, SCCH, Idea, UMA, Ideko, Tryis, IAS, Core, Farplas, Goimek, University of Wien, TU Dublin, Profactor Researchers: Mehmet Tuncel, Gozde Unal, N. Kemal Ure Dates: 2021-Present
Teaming.AI is a project funded under ICT-38 call under H2020. Conosrtium spans 15 partners and 5 countries. The main objective of the project is to bring next generation of AI tools into smart manufacturing systems to increase adoption of AI methodologies across EU region. ITU is leading the machine learning efforts in the consortium, with a specific focus on developing novel deep learning and deep reinforcement learning methods for extracting knowledge graphs, building digital twins, performing anomaly detection and predictive maintenance.
Advanced Perception, Planning and Learning for Agile Swarm Robotics
Partners and Supporting Agencies: Havelsan, Researchers: Onur Akgun, Mustafa Erdem, Kamil Canberk Atik, Alican Saatci, Mehmetcan Kaymaz, Bugrahan Yamak, Akif Sagtekin, Enes Erdogan, Gozde Unal, N. Kemal Ure Dates: 2020-Present
Synopsis: This project focuses on development of novel SLAM, segmentation and reinforcement learning algorithms for coordination of agile aerial and ground vehicles.
Behavioral Cloning for Fighter Aircraft Pilots with Active Deep Learning
Partners and Supporting Agencies: TAI, Researchers: Abdullah Sadik Satir, Gulay Sever, Umut Demir, Mustafa Cagatay Sahin, N. Kemal Ure Dates: 2020-Present
Synopsis: This project focuses on development of deep neural network models for behavioral cloning of fighter aircraft maneuvers. In particular, we focus on active learning methods where the data is obtained from real pilots in an interactive manner.
Graphical Deep Reinforcement Learning for Multiagent Military Decision Making (HAMLE)
Partners and Supporting Agencies: Aselsan, Researchers: Tolga Ok, Emircan Kılıçkaya, Bengisu Guresti, Batuhan İnce, Kubilay Kağan Kömürcü, N. Kemal Ure, Dates: 2018-Present
Synopsis: In this project we develop novel deep reinforcement learning algorithms that exploit relational structure in multiagent decision making problems by using graphical convolutions and self attention mechanisms.
Social Media Engagement Prediction with Deep Learning
Partners and Supporting Agencies: Lisa AI, Researchers: Onur Atılgan, Tarık Demir, Orkan Yılmaz, N. Kemal Ure, Dates: 2017-Present
Synopsis: We develop deep learning models for predicting engagement of social media posts on Instagram, using multiple data sources such as pictures, audience features and user patterns.
Partners and Supporting Agencies: Eatron Technologies, Researchers: Anıl Ozturk, Mustafa Burak Gunel, Resul Dagdanov, Ekim Vural, Ferhat Yurdakul, Ugur Yavas, N. Kemal Ure, Dates: 2019-Present
Synopsis: This project develops novel deep reinforcement learning methods for high level tactical decision making for autonomous driving, as well as advanced deep learning methods for high performance computer vision and perception.
Probabilistic Trajectory Prediction and Weapon Allocation for Structured and Coordinated Attack Patterns
Partners and Supporting Agencies: Aselsan, Researchers: Abdullah Sadık Satır, Gulay Sever, Umut Demir, N. Kemal Ure Dates: 2019-Present
Synopsis: In this project we develop fast and efficient machine learning algorithms for predicting future maneuvers and positions of threats executing structured attack patterns.
Optimization and Machine Learning Methods for Swarm to Swarm Engagement Planning
Researchers: Samet Uzun, Behcet Acikmese, N. Kemal Ure Dates: 2019-Present
Synopsis: In this project we focus on developing novel probabilistic collaborative guidance methods and model predictive control laws for large scale swarm to swarm engagement problems.
Partners and Supporting Agencies: Arcelik, Researchers: Aydin Ayazandeh, Omid Abdoullahi, Kevser Keser, Behcet Ugur Toreyin, N. Kemal Ure, Dates: 2019-Present
Synopsis: This project focuses on development of novel model compression algorithms for executing real-time visual analytics on edge devices.
Partners and Supporting Agencies: Turkish Technic, Dates: 2019-Present
Synopsis: This project focuses on developing high performance guidance and control algorithm for unmanned cargo aircraft.
Partners and Supporting Agencies: Delta V, Researchers: Kamil Canberk Atık, Cansul Yıkılmaz, N. Kemal Ure, Dates: 2018-Present
Synopsis: This project focuses on high performance navigation filter design and sensor fusion algorithm development for launch vehicles.
Retail Sales Prediction with Deep Learning
Partners and Supporting Agencies: Migros, Dates: 2019-Present
Synopsis: We develop deep learning models for predicting future sales based on large scale data consisting of historical sales, demographic and geographic features.
Data Driven Optimization for Airline Operations
Partners and Supporting Agencies: GE Aviation Digital, Researchers: Özgün Altunkaya, Anıl Yıldız, Leila Ghasemzadeh, Ceren Eryılmaz, N. Kemal Ure, Dates: 2018-Present
Synopsis: This project focuses on utilizing machine learning and large-scale linear programming techniques for solving decision making problems in airline operations, such as crew management.
Multi UAV Surveillance and Track for Border Security Applications
Partners and Supporting Agencies: Milsoft, Dates: 2018-2019
This project focuses on development of reinforcement learning algorithms for high level coordination of multiple unmanned aerial vehicles.
Geometric and Algebraic Methods in Multiagent Deep Reinforcement Learning with Applications to Forest Firefighting
Partners and Supporting Agencies: European Comission, Researchers: Ahmed Farabi Tarhan, N. Kemal Ure, Dates: 2017-2019
Synopsis: Multiagent planning problems are ubiquitous in engineering, with examples ranging from multiple mobile robots used in planetary exploration to use of multiple unmanned aerial vehicles for forest firefighting. Many multiagent planning problems can be formulated as Markov Decision Processes and an approximate optimal solution can be computed with dynamic programming and reinforcement learning methods. However, these algorithms usually become computationally infeasible with the increasing number of agents. In recent years, using deep neural networks in conjunction with reinforcement learning algorithms proved to be a powerful method in solving such problems. In this project, we demonstrate applications of geometric deep learning (convolutional deep learning layers extended to non-Euclidean domains) to solution of large-scale multiagent planning problems.We also develop algebraic statistics based algrorithms for selecting exploration strategies that yield faster convergence compared to heuristic approaches.
High Accuracy Missile Guidance, Navigation and Control
Partners and Supporting Agencies: Aselsan MGEO, Researchers: Burak Yüksek, Mehmet Uğur Akçal, Batuhan Hoştaş, Anıl Yıldız, N. Kemal Ure, Dates: 2017-2018
Synopsis: This project focuses on development of a high accuracy missile guidance module, as well as development of nonlinear control and state estimation algorithms. As a secondary objective, project focuses on cooperative missile guidance algorithms for intercepting highly maneuverable targets.
Fault-Tolerant Adaptive Control of Fighter Aircraft with Deep Learning
Partners and Supporting Agencies: TAI, Researchers: Burak Yüksek, Batuhan Eroğlu, Mustafa Çağatay Şahin, N. Kemal Ure, Dates: 2017-2019
Synopsis: This project focuses on development of adaptive control algortihms that would enable safe recovery and autonomous landing for fighter aircraft that experience severe actuator and engine failures. We use deep learning based models for assesing the failure state of the aircraft, which enable rapid convergence of the adaptive control components, which in turn enable fast recovery.
Decision Making Under Uncertainty and Deep Learning for Self-Driving Cars
Partners and Supporting Agencies: AVL, Researchers: Majid Moghadam, Yunus Biçer, Yaser Moazzen, Sahand Vahidnia, N. Kemal Ure, Dates: 2018
Synopsis: This project focuses on building planning under uncertainty algorithms for self-driving in dynamic environments. In particular, we develop efficient deep learning models for computer vision, decision making and driver intention prediction using both high fidelity simulators and data obtained from real driving tests. As a secondary objective, project focuses on multi vehicle network coordination and simultaneous localization and mapping using deep learning algorithms.
Structural Health Monitoring and Failure Prediction with Deep Learning
Partners and Supporting Agencies: GE Power, Researchers: Mahtab Khani, Leila Ghasemzadeh, Sahand Vahidnia, N. Kemal Ure, Dates: 2018
Synopsis: This project focuses on building deep learning based classification and prediction models for detecting structural failures in wind turbine components. The algorithms are deployed on robotic arms that perform autonomous inspection on the components.
Maneuver Planning and Control for Agile Maneuvering Aircraft
Partners and Supporting Agencies: TUBITAK, Researchers: Majid Moghadam, Mehmet Uğur Akçal, Batuhan Hoştaş, Anıl Yıldız, N. Kemal Ure, Dates: 2016-2018
Synopsis: The aim of this project is to develop control systems that would enable agile maneuvering for fixed and rotary wing Unmanned Aerial Vehicles (UAVs) along with the development of flight envelope protection and maneuver planning algorithms. The first and foremost objectives of the project are to expand the autonomously executable maneuver set of the military oriented UAVs and optionally piloted aircraft systems, prevent loss of control situations induced by the execution of these agile maneuvers as and hence increase the lifespan of UAVs operating in adversarial conditions. In this context, the project’s expected outcomes are algorithms that would enable the UAV to compute agile maneuvers for defending itself from threats (such as missiles and other aerial adversaries) and counter-attacking them, tracking these reference maneuvers with high accuracy and return autonomously safe flight conditions if these agile maneuvers force the aircraft the violate the flight envelope due to high angle of attack and high translational speed. The developed algorithms are verified with simulation studies on six degrees of freedom nonlinear UAV models for various different threat scenarios.
Threat Modeling and Scenario Generation for Air Missile Defense Systems
Partners and Supporting Agencies: Aselsan SST, Researchers: Burak Yüksek, Mehmet Uğur Akçal, Hakkı Karakaş, Gökhan İnalhan, Onur Tuncer, Adil Yükselen, Hayri Acar, Zahit Mecitoğlu, N. Kemal Üre, Dates: 2015-2016
Synopsis: This project developed a diverser set of 3-DOF threat models and an attack pattern generation tool for testing the efficiency of air missile defense systems.
Health Aware Planning for Multiagent Persistent Missions
Partners and Supporting Agencies: Boeing Research and Technology, Researchers: N. Kemal Ure, Jonathan P. How, Girish Chowdhary, Yu Fan Chen, Tuna Toksoz, Josh Redding, Matthew Vavrina, John Vian Dates: 2010-2015
Synopsis: The goal of this work is to create methods to solve online multiagent planning problems under constraints such as, uncertain mission dynamics, communication requirements and time varying agent capabilities. The work combines techniques from approximate dynamic programming, model based reinforcement learning and learning focused adaptive control.
Autonomous Control of Unmanned Combat Air Vehicles: Design of a Multimodal Control and Flight Planning Framework for Agile Maneuvering
Researchers: N. Kemal Ure, Emre Koyuncu, Gokhan Inalhan, Dates: 2006-2010
Synopsis: This project studies the problem of generation and control of maneuver profiles for Unmanned Combat Air Vehicles (UCAV) with aggressive maneuvering capability. A motion description framework is developed, where complex maneuvers are decomposed into series of sub-maneuvers (modes) and associated parameters. Main advantage of framework is allowance of the feasibility and control problems to be treated for each sub-maneuver individually instead of whole maneuver, thus reducing the complexity of the problem which can also be modeled in the language of Hybrid Dynamical Systems. For maneuver generation, agility metrics and mode compatibility relationships are used to create a feasible maneuver generation algorithm. For low level controls, each mode is assigned a Higher Order Sliding Mode Controller, where whole maneuver is controlled via switching between modes and controllers. Overall generation and control algorithms are tested on agile combat maneuver examples.