Mobile Robots in Human Environments: towards safe, comfortable and natural navigation

Traditionally, robots have been assistant machines in factories and a ubiquitous part of science fiction movies. But within the last decade the robots have started to emerge in everyday human environments. Today they are in our everyday environment in the shape of, for example, vacuum cleaners, lawn mowers, toy pets, or as assisting technologies for care giving. If we want robots to be an even larger and more integrated part of our every- day environments, they need to become more intelligent, and behave safe and natural to the humans in the environment. This thesis deals with making intelligent mobile robotic devices capable of being a more natural and sociable actor in a human environment. More specific the emphasis is on safe and natural motion and navigation issues.
First part of the work focus on developing a robotic system, which estimates human interest in interacting with the robot. This information is then used to navigate appro- priately in the environment around a person. The system consists of three parts; finding the person’s state information (position, velocity and orientation) relative to the robot, determining if the person wants to interact with the robot, and finally human-aware navi- gation that respects the persons social zones and interest in interaction. From laser range scanner measurements, a new Kalman filter based method is used to infer the person’s state information. Secondly, the robot adaptively learns to estimate if a person seeks to interact with the robot. This is done using a Case-Based Reasoning System (CBR), which analyses current behaviour patterns while comparing these to experiences and outcomes from previous human encounters. The motion of the robot is controlled based on adaptive potential functions. The potential functions are adjusted in accordance with the current interest in interaction and such that the person’s social spaces are respected. The opera- tion of the system is evaluated in an open hall setting at the university. It is demonstrated, that the robot is able to learn where to position itself, and is capable of adapting to a changing environment. It is then shown that the developed framework can be used in other applications as well. The framework is transformed into a robotic game that, for example, can be used to motivate elderly people to a regular amount of exercise, which then keep them in better physical shape.
As well as being able to navigate safely around one person, the robots must also be able to navigate in environments with more people. This can be environments such as pedestrian streets, hospital corridors, train stations or airports. The developed human-aware navigation strategy is enhanced to formulate the problem as planning a minimal cost trajectory through a potential field, defined from the perceived position and motion of persons in the environment. A Rapidly-exploring Random Tree (RRT) algorithm is proposed as a solution to the planning problem, and a new method for selecting the best trajectory in the RRT, according to the cost of traversing a potential field, is presented. The RRT expansion is enhanced to account for the kinodynamic robot constraints by using a robot motion model and a controller to add a reachable vertex to the tree. Instead of executing a whole trajectory, when planned, the algorithm uses a Model Predictive Control (MPC) approach, where only a short segment of the trajectory is executed while a new iteration of the RRT is computed. The planning algorithm is demonstrated in a simulated pedestrian street environment.
One problem with sampling based methods for trajectory planning, is the high number of samples, which are necessary to plan a good trajectory. Therefore, a method, which decrease the computation time for an RRT, such that more vertices can be added in the same amount of time to generate better trajectories, is developed. The algorithm is based on subdividing the configuration space into boxes, where only specific boxes needs to be searched to find the nearest neighbour. The result is an algorithm that can provide better trajectories within a given time period, or alternatively compute trajectories faster. In simulation the algorithm is verified for a simple RRT implementation and in the more specific case where the robot has to plan a path through an environment with moving people.