Consider a science mission in which a marine robot is to follow and observe the behavior of a rare type of fish that feed on coral reefs. During its journey, the robot must:
- not damage the endangered coral reefs,
- not lose the sight of the rare fish,
- have the ability to collect accurate location and environment data, and
- maintain a sufficient coverage of the area of interest in order to increase the probability of observing any interesting phenomena. Such scenarios pose extremely challenging problems for autonomy, involving decision making, trajectory planning, and tracking in a highly uncertain environment. In addition, the robot dynamics are nonlinear with uncertain parameters, perception data are noisy, communication with other platforms is expensive, and GPS localization is virtually nonexistent in underwater domains. The mission requirements also not completely specified. Even though a high-level descriptions may be available, many details can be left unspecified.
Motivated by these issues, we aim to explore how human interaction with vehicles can be used to efficiently learn task specifications and autonomous behaviors in uncertain domains with incomplete/uncertain information. As part of this work, we develop new algorithmic strategies for efficiently programming and improving underwater vehicle autonomy in practical applications. Another aim of the project is to build a software stack and hardware testbed to support underwater vehicle research as well as developing new algorithmic strategies for efficiently programming and improving underwater vehicle autonomy in practical applications.
K. Watanabe et al., “Timed Partial Order Inference Algorithm,” in International Conference on Automated Planning and Scheduling (ICAPS), Prague, Czech Republic, 2023. (to appear)
G. O. Berger, M. Narasimhamurthy, K. Watanabe, M. Lahijanian, and S. Sankaranarayanan, “An Algorithm for Learning Switched Linear Dynamics from Data,” in Advances in Neural Information Processing Systems (NeurIPS), New Orleans, Louisiana, USA, 2022.
K. Watanabe, N. Renninger, S. Sankaranarayanan, and M. Lahijanian, “Task Learning with Preferences for Planning with Safety Constraints,” in Workshop on Accessibility of Robot Programming and the Work of the Future, 2021.
K. Watanabe, N. Renninger, S. Sankaranarayanan, and M. Lahijanian, “Probabilistic Specification Learning for Planning with Safety Constraints,” in IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021.