RI: Small: Information-theoretic Multiagent Paths for Anticipatory Control of Tasks (IMPACT)

This project promotes the scientific and engineering value of intelligent navigation systems by finding the best routes of autonomous robots based on the desired level of exploration, risk, and energy constraints. High-performance onboard computing enables fundamentally new computational research on cooperative navigation between unmanned aerial and ground vehicles in real-time but also raises new challenges in using acquired, but imperfect information of the system. Robots analyze images for autonomous driving feature detection and assist scientists by selectively collecting data without interrupting drives. This project discovers informative paths during the journey, updated as information about the terrain is discovered. The simulation will provide valuable insights into utilizing the map. Understanding how information can be learned throughout navigation will produce a guide to robotics planners, and offer substantial benefits to society in improved choice modeling. This research will have a positive impact in emergency situations when some of the road networks are disconnected. Future autonomous vehicle driving will incorporate energy efficiency by considering the tradeoff between energy efficiency and congestion.

This project introduces a set of novel techniques for probabilistic information gain by predicting potential speed classification of future arrival locations to improve rover productivity by extending travel distance, allowing more time for non-driving activities, and reducing the required solar cell area. Route planning of Mars robots depends on a traversability estimate, based on orbital imagery, which varies in confidence level at different locations. Typically, a visual inspection of images reveals a finite number of distinctive terrain units, where the traversability is likely near-homogeneous within each unit. Therefore, once a robot visits a part of a terrain unit and images it, the uncertainty in traversability of the other parts of the same unit is reduced. This reasoning results in the concept of information-theoretic route planning: visiting high-uncertainty areas at the early stage of a mission to resolve uncertainty (i.e., information gain) and benefit the future route planning. However, such exploratory behavior is justified only when the benefit from uncertainty reduction exceeds the cost of exploration. Particular considerations include 1) a sequential information gain from single observation against multiple observations, 2) a mixture of information gain in multiple univariate probability distributions against the multi-variate setting, and 3) implementation to energy-aware planning with information gain. When a robot travels through a grid map, information can be gained by visiting unclassified or uncertainly classified cells, observing the condition in those cells, and estimating the entropy in other cells. Each agent updates its path plan every time it moves to a new grid cell. By sharing information about the state of the grid cells, each agent helps to define the optimal parameters to be used in other agents' utility functions. If an identical cell is visited by another agent and found to be in the same state as the original cell of that type, then all agents have confirmation that the assumption that these cells are correlated is more likely to be true. The degree of uncertainty in the map depends on the time-dynamics of the agents' visits, information obtained through satellite imagery, and the upper and lower bound of travel times.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.