Author:

Abstract:

Challenge. The demographics of society are changing because of an ageing population. As a consequence, disabilities related to ageing are becoming more prevalent. Further, fewer young people are available to attend the elderly or disabled. Such evolution is problematic since disabilities often involve the loss of mobility and affected individuals depend on others for help. Impaired mobility can also lead to isolation. Both effects combined potentially reduce a person's well-being and self-esteem. Besides the important personal aspect, loss of mobility and dependency can also have a grave economic impact on society as a whole.Fact. Mobility enhancing technology, such as electrical wheelchairs, addresses the demographic challenge by helping people to regain their independence. However, electrical wheelchairs are difficult to control and dangerous accidents with the environment and other people are rife. Therefore, many potential users are not prescribed an electrical wheelchair if they lack the necessary cognitive or physical skills. Even able users find it difficult and tiresome to navigate a wheelchair safely. Opportunity. The past decade has witnessed tremendous advances in mobile robotics. Autonomous cars were racing through the dessert or navigating in a city. Mobile robotics is also finding its way to mainstream products, such as vacuum cleaners or lawn mowers. Hence, mobile robotics is an ideal technology for helping the elderly and disabled to navigate safely in almost any environment. Challenge. The elderly and disabled are a very heterogeneous user group, with different abilities and needs. The challenge is to develop customised robots, which can work together effectively with such a heterogeneous population, but still remain easy and intuitive to control.Requirements. Such a heterogeneous user group requires a general methodology that can adapt navigation assistance to each individual. In order to avoid frustrating its user, the robot should attempt to recognise its user's navigation plan in a way that considers the user's driving abilities. Furthermore, the robot should provide assistance only when the user requires it. Users should be encouraged to participate in navigation as far as possible to remain mentally and physically fit.Methodology. In this thesis we propose a probabilistic framework to recognise the user's navigation plans out of a set of local trajectories. In order to increase robustness, this framework considers the uncertainty when recognising user plans. It fuses past driving information with the user's specific driving style in order to estimate a posterior probability over user plans. Further, these estimated plans are the basis to share the control over the wheelchair according to the user's abilities and needs. The robot ensures the user's safety because it reasons about collision-free trajectories that are also physically executable,i.e. kinematically and dynamically feasible. The first contribution of this thesis is a probabilistic formulation of human-robot navigation as a Dynamic Bayesian Network (DBN). In order to achieve effective collaboration, the robot estimates the user's plan out of a set of local trajectories. During this process, the robot must propagate past driving information to the present in absence of global references. Therefore, we propose a solution to local plan recognition within the human-robot DBN through soft probabilistic constraints. These constraints model the uncertain information a potential user plan provides about the actually executed trajectory. We summarise such information in a distance likelihood, which we combine with other observations in order to derive the parameters of the human-robot DBN with the Expectation Maximisation algorithm. We illustrate the validity of the approach on a simulated wheelchair with a discrete interface as well as on a real wheelchair operated through a continuous joystick. The second contribution is a set of probabilistic user models of driving. We introduce first an analytical user model of a user who cannot steer left. The second user modelling approach calibrates the user's driving style from real data through a general probabilistic machine learning technique called Gaussian Process Regression. We justify the selection of this technique with an analysis of user modelling requirements, and address how specific choices of the model fulfil them. In order to demonstrate the validity of the approach we analyse the driving behaviour of a user diagnosed with spastic-quadriplegia and recognise her navigation plans.The third contribution is a framework for sharing the control of the wheelchair according to the estimated plans of its user. This framework is based on the Dynamic Window Approach. We propose three shared control algorithms where the user increasingly cedes more control to the robot. The first two shared control schemes correspond to standard navigation algorithms, namely collision-avoidance and obstacle-avoidance, which are special cases within our general framework. The third algorithm, Maximum A Posteriori (MAP) shared control, relies on the estimated plan of the user to navigate the wheelchair. We demonstrate the potential of user-adaptive shared control for a user who cannot steer left, but thanks to navigation assistance can reach goals at her left when she intends it. This experiment reveals that navigation assistance benefits from knowing the driving style of the wheelchair user, and that bi-directional communication in case of disagreement is necessary to share the control over the robot effectively.