When to Localize? A POMDP Approach

TL;DR

This study proposes a method that helps a robot determine “when to localize” to 1) minimize such actions and 2) not exceed the probability of failure (such as surfacing within high-traffic shipping lanes).

Abstract

Robots often localize to lower navigational errors and facilitate downstream, high-level tasks. However, a robot may want to selectively localize when localization is costly (such as with resource-constrained robots) or inefficient (for example, submersibles that need to surface), especially when navigating in environments with variable numbers of hazards such as obstacles and shipping lanes. In this study, we propose a method that helps a robot determine ``when to localize'' to 1) minimize such actions and 2) not exceed the probability of failure (such as surfacing within high-traffic shipping lanes). We formulate our method as a Constrained Partially Observable Markov Decision Process and use the Cost-Constrained POMCP solver to plan the robot's actions. The solver simulates failure probabilities to decide if a robot moves to its goal or localizes to prevent failure. We performed numerical experiments with multiple baselines.

Figures (6)

• Figure 1: An underwater vehicle follows a path to perform a task (like searching for an aircraft's black box) and must choose when to localize, which requires surfacing and poses a collision risk in shipping lanes. Continuous localization (bottom path) is inefficient yet safe. Selective localization (top path) is more efficient and desirable since the vehicle searches underwater longer. But deciding when to localize to avoid hazards and stay on the path is more challenging.
• Figure 2: This figure illustrates our training environment (ENV-TRAINING), a $30\times30$ 2D grid world we used to determine the parameters for our online algorithms (shown in Table \ref{['table:op_parameters']}). The salmon-colored rectangles denote shipping lanes where boats sail while the AUV navigates underwater.
• Figure 3: This figure depicts a screenshot of ENV-TUNNEL, one of our evaluation environments. The environment contains obstacles in brown, localize hazards (such as shipping lanes) in red, the (yellow) start and (green) goal positions, and the pre-defined path the AUV should follow.
• Figure 4: This figure illustrates our real-world evaluation environment called ENV-STT. The image shows a grid overlaying a Google Maps screenshot of the southwestern portion of St. Thomas, U.S. Virgin Islands. The green cells represent obstacles (land masses), the grey cells denote open water, and the white cells represent the AUV's pre-defined path. The AUV starts at the University of the Virgin Islands' (UVI) Marine Center and ends at Lindbergh Bay Beach.
• Figure 5: The Physical Failure Rates (top) and Cumulative Percentage of Collisions (bottom). This figure depicts the results for the ENV-TRAINING (left), ENV-TUNNEL (middle), and ENV-STT (right) environments. Here, the failure rates refer to the number of times the robot left the map or collided with an obstacle while moving or localizing. Each graph shows the mean and standard deviation for each algorithm. Due to POMCP's poor performance, we clipped the cumulative percentages (bottom graphs) to 100%. Finally, the red dashed horizontal lines in the bottom graphs show the collision threshold of 10%.