THINKING FOR THEMSELVES: Autonomous robots plan their own actions in complex, dynamic environments By Jonathan P. How, Emilio Frazzoli, and Brian Williams Automation has the potential to greatly improve the performance and capabilities of future robotic vehicles, such as cars and airplanes.
The reader might be familiar with autonomous vehicles, such as the Roomba vacuum cleaning robot, that perform fairly simple tasks in constrained environments — but the
goals of this work are to develop much smarter robots that can operate in, and adapt to, the dynamic and complex world familiar to humans. This includes examples such an autonomous car that optimizes its route to a pick-up point to avoid traffic tie-ups reported by other vehicles, while designing motion paths that avoid potentially erratic maneuvers by pedestrians, cyclists, and other cars. It could also include a team of UAVs collaborating to choose the best locations to sample the atmosphere to better predict a storm’s path. From choosing which actions to take, to deciding which maneuvers to execute to perform those actions, planning is a crucial part of optimizing the solution to the fundamental questions of “what, where, when, how?” Of course, to operate in a dynamic world, this planning must be fast enough that it can be performed online, in real-time. The result is similar to classical feedback control, in which sensors provide information about the motion of the system, and then use the information to compute corrective inputs. But, autonomy extends that framework to include both continuous and discrete decisions, such as activity planning, compliance to rules of behavior, and trajectory design. There are numerous challenges to achieving these goals, such as developing planning algorithms that are both computationally tractable and communication-efficient, adapting the autonomous
THINKING FOR THEMSELVES: Autonomous Robots Plan Their Own Actions in Complex, Dynamic Environments