top of page

Esen Yel

I am currently a Ph.D. student in System Engineering at the University of Virginia. I work in Autonomous Mobile Robots Lab and Link Lab under the supervision of Prof. Nicola Bezzo. Before joining AMR Lab, I obtained my MS and BS degrees in Electrical and Electronics Engineering at Bogazici University, Istanbul in 2016 and 2014 respectively. 

My research focuses on safety-guaranteed motion planning, reachability analysis, assured autonomy, run-time monitoring and recovery, self-triggered scheduling and control of autonomous aerial vehicles. 



Reachability-based Self/Event-triggered Scheduling and Replanning

Modern unmanned aerial vehicles (UAVs) rely on high-frequency periodic sensor measurements for accurate and safe autonomous operations in static or dynamic environments with uncertainties. However, periodic sensor measurements can be computationally consuming, and they are not always necessary, especially when the vehicle is operating in obstacle-free environments.

Reachability analysis is leveraged to minimize the number of sensor monitoring operations while guaranteeing safety (i.e., the vehicle is not going to collide with an obstacle) and liveness (i.e., the vehicle will follow its trajectory closely) in both static and dynamic environments under external disturbances and noises. More details about the approach and the simulation and experiment results can be found in our papers. [AHS '17], [ICRA '18], [JINT '20]


GP-based Fast Runtime Monitoring and Recovery

In real-world applications, the UAVs may not always be able to monitor their state information due to sensor failures, signal occlusions, and communication problems (e.g., losing GPS signal while moving in a city around tall buildings). Model-based reachability analysis techniques are powerful to provide safety guarantees, but they are computationally expensive.

Gaussian Process regression is leveraged to perform fast reachability analysis, and to schedule when the system needs to monitor, replan and recover to guarantee safety. More details can be found in our paper. [IROS '19]

website_fig (1).png

Assured Runtime Monitoring and Planning

Autonomous systems operating in uncertain environments under the effects of disturbances and noises can reach unsafe states even while using well-tuned controllers. To make safety predictions on such systems during motion planning operations, neural networks (NNs) are trained offline using reachability analysis. If the designed trajectory is predicted to be unsafe, trajectories are replanned until a safe one is obtained. To guarantee safety, the trained neural networks are verified using Verisig. In the case of unverified NNs, Verisig output is used to retrain NNs more conservatively. With this approach, the reachable set computation is limited to the offline stage, and safety-assured trajectories are designed at runtime in both known and unknown environments. [RAM '20]


Runtime Planning, Learning, and Recovery under Unforeseen Disturbances

Autonomous systems are typically designed and trained to work under certain system and environmental conditions. However, in real-world applications, the system may face unexpected situations such as component failures, variation in payload distribution or changes in model dynamics at runtime. Such factors could lead the system to undesired conditions if they are not properly taken into consideration.

To provide safety under unforeseen distrurbances at runtime, a fast online planning, learning, and recovery approach is introduced. By leveraging Gaussian Process regression theory in which a model is continuously trained and adapted using data collected during the autonomous operation, the disturbance and its effects are predicted at runtime to plan safe trajectories. This approach is validated on simulations and experiments on a UAV carrying an unknown payload outside of the training bounds. [IROS '20]



When I am not working, I enjoy yoga, hiking, traveling, collecting vintage photographs, watching independent movies, and reading.

bottom of page