Autonomous systems may need to operate under different conditions than they have been designed for (e.g., unforeseen faults, different environment settings, etc.). To deal with this problem, the systems need to be able to adapt their models to new situations without the need for large amounts of data. Our work has focused on leveraging meta-learning and Bayesian techniques to provide adaptation to unforeseen conditions.
- A. Yildiz, E. Yel, A. Corso, K. Wray, S. Witwicki and M. Kochenderfer, “Experience filter: Transferring past experiences to unseen tasks or environments”, IEEE Intelligent Vehicles Symposium (IV) 2023 PDF
- E. Yel, Shijie Gao, N. Bezzo, ”Meta-Learning-based Proactive Online Planning for UAVs under Degraded Conditions”, (*equal contribution), Robotics and Automation Letters (RA-L), 2022 PDF
- E. Yel, N. Bezzo, ”A Meta-Learning-based Trajectory Tracking Framework for UAVs under Degraded Conditions” IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2021 PDF
- E. Yel, N. Bezzo, ”GP-based Runtime Planning, Learning, and Recovery for Safe UAV Operations under Unforeseen Disturbances” IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2020 PDF
Autonomous systems can be subject to uncertainties due to external disturbances or due to the use of internal learning-based components. Since these systems are mainly used in safety-critical settings, it is important to assess the safety of the system before and during its deployment. Our research uses reachability analysis and neural network verification techniques to provide safety-critical autonomous systems with such capabilities.
- N. Rober, S. M. Katz, C. Sidrane, E. Yel, M. Everett, M. J. Kochenderfer, and J. P. How. “Backward reachability analysis of neural feedback loops: Techniques for linear and nonlinear systems”, IEEE Open Journal of Control Systems, 2023 (Early Access) PDF
- E. Yel, T. Carpenter, C. di Franco, R. Ivanov, Y. Kantaros, I. Lee, J. Weimer, N. Bezzo, ”Assured Run-time Monitoring and Planning: Towards Verification of Deep Neural Networks for Safe Autonomous Operations”, Robotics and Automation Magazine, Special Issue on Deep Learning and Machine Learning in Robotics, 2020 PDF
For safe navigation and planning, it is important for autonomous systems to accurately predict how other agents in their environments are going to move over time. Our research uses various machine-learning techniques ranging from Gaussian Processes to deep neural networks to enable systems to predict the future states of their environment so that they can proactively plan safe trajectories. These techniques have been applied in various application areas such as aerial robotics, autonomous driving, and underwater robotics.
- M. Toyungyernsub, E. Yel, J.Li, M. Kochenderfer, “Dynamics-Aware Spatiotemporal Occupancy Prediction in Urban Environments”, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2022 PDF
- M. Cleaveland, E. Yel, Y. Kantaros, I. Lee, N. Bezzo, “Learning Enabled Fast Planning and Control in Dynamic Environments with Intermittent Information”, IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) 2022 PDF
- L. Kruse, E. Yel, R. Senanayake, M. Kochenderfer, “Uncertainty-Aware Online Merge Planning with Learned Driver Behavior”, IEEE International Conference on Intelligent Transportation Systems (ITSC), 2022 PDF
- E. Yel and N. Bezzo, “Fast Run-time Monitoring, Replanning, and Recovery for Safe Autonomous System Operations” 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Macau, China, pp. 1661-1667. PDF
Reachability-based Self/Event-triggered Scheduling and Replanning
High-frequency periodic sensor measurements are usually expected for a traditional system to safely navigate in environments occupied with obstacles. However, monitoring the sensors and processing the sensor measurements at high frequency are computationally consuming, and often not necessary especially when the vehicle is operating in obstacle-free environments. Our work uses reachability analysis together with self-triggered scheduling to minimize the computation associated with sensor monitoring while guaranteeing safety.
- E. Yel, Tony X. Lin, N. Bezzo, ”Computation-Aware Adaptive Planning and Scheduling for Safe Unmanned Airborne Operations” Journal of Intelligent and Robotic Systems, 2020 PDF
- E. Yel, T. X. Lin and N. Bezzo, ”Self-triggered Adaptive Planning and Scheduling of UAV Operations,” IEEE International Conference on Robotics and Automation (ICRA), Brisbane, 2018 PDF
- E. Yel, T. X. Lin and N. Bezzo, ”Reachability-based self-triggered scheduling and replanning of UAV operations,” NASA/ESA Conference on Adaptive Hardware and Systems (AHS), Pasadena, CA, 2017, pp. 221-228. PDF