Case-based mobile manipulation

The ability to manipulate the environment is one of the primary skills that autonomous mobile service robots are expected to have, considering that the daily lives of humans heavily rely on this skill. There are various ways for a mobile robot to perform manipulation, the exact form of which is determined by the requirements of the task and the constraints imposed by the physical properties of the environment, the object, and the robot itself. Anecdotal evidence suggests that humans mostly reuse their manipulation experiences, acquired through interaction and observation, especially in recurring everyday manipulation tasks, both in prehensile and non-prehensile manipulation contexts. With this motivation, this thesis contributes a case-based approach to achieving practical and efficient mobile manipulation through the utilization of past experience, stored as object-specific, distinct, and potentially probabilistic cases. In scenarios where the objects of interest can be grasped, we observe that humans tend to reach for the targets (i.e. the object of interest or the destination) roughly and as directly as possible while exploiting a small set of target-specific critical moves within their close proximity to guarantee successful completion of the task. Inspired by the critical yet recurring nature of these target-specific relative moves, we let the robot memorize and store a small number of them as cases in its case base, and reuse them whenever possible to guide manipulation planning and execution. When combined with sampling-based generative planners, this guidance helps reduce planning time by deliberately biasing the planning process towards the feasible cases to achieve the rough and direct approach effect. Additionally, monitoring the execution while reiterating the reached cases improves task success rate. Our experiments show that this complementary combination of the already available partial plans and executions with the ones generated from scratch yields to fast, reliable, and repeatable solutions. When the objects of interest are too heavy or bulky, or a grasping mechanism is not available, it may still be possible to perform manipulation through pushing. However, in realistic mobile push-manipulation scenarios, the complexity of the potential physical interactions between the robot, the objects, and the environment is vast. This makes it non-trivial to write generic analytical motion and interaction models or tune the parameters of physics engines for every robot, object, and environment combination. In order to handle this problem flexibly, we introduce a case-based push-manipulation framework that enables the robot to acquire, through interaction and observation, a set of discrete, experimental, probabilistic motion models (i.e. probabilistic cases) for pushable passively-mobile real world objects. These probabilistic cases are then used as building blocks for constructing safe and achievable push plans to navigate the object of interest to the desired goal pose as well as to potentially push the movable obstacles out of the way in cluttered task environments. Additionally, incremental acquisition and tuning of the probabilistic cases allows the robot to adapt to the changes in the environment, such as increased mass due to loading of the object of interest for transportation purposes. The purely interaction and observation driven nature of our method makes it robot, object, and environment (real or simulated) independent, as we demonstrate through extensive testing and experimentation. We also verify the validity of our push-manipulation method in preliminary real world tests.