|Speaker :||Sara Ayoubi|
|Time:||3:00 pm - 4:00 pm|
Coordinated path planning and motion control are crucial for collision-free navigation of a fleet of robots that share a space and potentially contend for the same paths and passages. Many existing navigation software packages are designed for single-robot navigation and treat other robots as any other observed obstacle, but this strategy alone is not sufficient for the system to work reliably. The ability to use the existing single-robot stacks without modification while also achieving coordination has high practical value. This talk will present a system that achieves coordination without modifying the path planner or motion controller. The only requirement, which exists already in most navigation stacks and is otherwise easy to implement using existing interfaces, is that the navigation stack provide interfaces for injecting dynamic prohibition regions into the cost map and for requesting that the robot temporarily stop its motion. The system, which we call PA-CODA, observes the robots’ intents and uses a learned hidden Markov model to anticipate collisions, which are then avoided by manipulating the cost map or temporarily stopping one or more robots. We ran a series of experiments with real robots to demonstrate PA-CODA’s ability to successfully detect and avoid collisions.