OREGON STATE UNIVERSITY
Open search box
Autonomous approaches are effective, but do not work in every environment. Autonomy suffers in situations when computer vision fails, or when the learned task is too dissimilar to the current task. Knowing the risk of using an autonomous or alternative approach is important when deciding how to perform a task. In this work, we introduce A3P , a risk-aware task-level reinforcement learning algorithm. A3P represents a task-level state machine as a POMDP. In this POMDP tasks are states and approaches are actions. Failures are represented in the learned solution as additional state-action pairs, allowing the user to make more informed decisions when choosing between different approaches.
Currently all work on A3P has been in simulation. Further validation is required, and we will apply A3P to a real-world robotic navgiation and task scenario.