My research centers on the design of innovative machine learning algorithms for practical scenarios with limited datasets and/or constraints on data collection resources. More specifically, I below describe some of my ongoing interdisciplinary research projects.

Prediction of Unseen Samples

I am interested in creating anomaly detectors using “only” normal, typical observations, which are more accessible and cost-effective to obtain. For instance, in the context of automating fruit health monitoring from image data, it can be challenging to collect sufficient samples of unhealthy fruits. Even if you have some examples, completely novel types of disease could arise later, posing new challenges for detection. To overcome this, I develop methods to leverage the normal samples to synthesize hypothetical but realistic examples of anomalies so that these synthetic anomalies can serve as valuable training data for accurate anomaly detection. In other words, my research aims to enhance monitoring systems by bridging the gap between limited anomaly data and the need for comprehensive anomaly detection solutions. Specifically, I have shown novel techniques that utilize self-supervised learning and generative adversarial networks to tackle practical scenarios like fruit health monitoring and ant behavior analysis.

References:

• Self-supervised Representation Learning for Reliable Robotic Monitoring of Fruit Anomalies. In: Proceedings of the 2022 IEEE International Conference on Robotics and Automation (ICRA 2022). [Data] [Code].
• Beyond Tracking: Using Deep Learning to Discover Novel Interactions in Biological Swarms. Journal of Artificial Life and Robotics (AROB). [Data].
• Identification of Abnormal States in Videos of Ants Undergoing Social Phase Change. In: Proceedings of the 35th AAAI Conference on Artificial Intelligence (AAAI 2021). [Data] [Code].

Active, Informative Sampling

For environmental monitoring, the regular assessment of physical systems is crucial to detect and report changes in air or soil quality, fruit maturity, and more. To achieve this, embodied agents can serve as valuable tools, but their motions and paths for sensing must be optimized to maximize information gathering while considering operational limitations, such as limited battery life. In this project, specifically, I am focused on developing novel reinforcement learning algorithms to intelligently guide these agents in selecting the next optimal motions based on current observations. The objective is to collect highly informative data for building accurate models of the environment. Through my research, I have achieved promising results, including the development of a path planning method for air-quality mapping. In addition, I have created a publicly available agricultural dataset specifically designed to advance the study of active perception in fruit monitoring, considering challenging occlusion scenarios.

References:

• DAVIS-Ag: A Synthetic Plant Dataset for Developing Domain-Inspired Active Vision in Agricultural Robots. arXiv, 2023. [Data]

• Adaptive Selection of Informative Path Planning Strategies via Reinforcement Learning. In: Proceedings of the 10th European Conference on Mobile Robots (ECMR 2021).