Machine Learning Blog | ML@CMU | Carnegie Mellon University
Active Search (Active Sensing) in robotics refers to the problem of searching for objects of interest in an unknown environment by making data-collection decisions. Active Search has many applications including the search and rescue of human survivors following disasters, detecting gas leaks or locating and preventing animal poachers. Our focus in this work is to perform active search using aerial or ground robots. We intend to use multiple robots to parallelize and speed up the search.
To perform active search using multiple robots, we need a technique to coordinate the actions of all robots and help them decide where to sense at every point in time. One way to approach this problem is to have a center plan and coordinate actions of all robots. This approach requires the center to maintain a synchronous and constant communication channel to all robots. However, maintaining constant communication is very difficult in certain applications, such as surveillance, exploration of unknown environments and search-and-rescue. For example, moving in an unknown or cluttered environment, it is very likely for robots to get trapped and temporarily lose their connection to the center [1]. As a result, a central coordinator that expects synchronicity from all robots at all times is not feasible as any robot failure or communication delay could disrupt the entire process. To clarify, we can still assume the robots are communicating with each other in a decentralized manner to share their observations. It is just that their communication links can be unreliable at times.
With this motivation in mind, we are interested in developing a decentralized multi-robot algorithm that helps robots independently decide where to go to sense at every point in time. This is typically viewed as the problem of intelligently choosing a waypoint to navigate towards. Each robot can use any available information shared from other robots’ sensed locations to decide which waypoint to navigate to next. Since communication between robots can be unreliable, our proposed algorithm must be able to use only partial information shared among robots and never depend on full information to make decisions.
