Human-machine cooperative telerobotics using uncertain sensor and model data

Recent years have seen an increase in the number and complexity of tasks in remote, dangerous, or inaccessible areas, such as nuclear, underwater, space, or microscopic environments requiring remote manipulation and sensing technologies. Due to the limitations of machine intelligence in unstructured or dynamic environments, humans are often an integral part of the control loop. This dissertation addresses the issues of combining human control with machine assistance by using available sensor information. The approach combines differential geometry techniques of position and velocity mapping of commands from the master to the slave manipulators and selection of parameters in these strategies with a probabilistic representation of task and environment. Operator commands are merely augmented or diminished, rather than superseded, based on the confidence in the utilized sensor information. In this way, even inaccurate sensor data may be used with some benefit. Preliminary experiments, including surface impact and Fitts task, were used to demonstrate the efficacy of the algorithms and provided a simple example of the broader framework of assistance. Further experiments concentrated on assistance to practical tasks of interest by the DOE involving radioactive waste tank cleanup. The application of the assistance strategy resulted in clear improvements to execution efficiency, fatigue reduction, and safety.