A data-driven hierarchical framework for planning, navigation, and control of uncertain systems: applications to miniature legged robots

Tanner, Herbert G.Poulakakis, IoannisPerforming navigation with state-of-the-art mobile robots in real-world settings is challenging because of, among other reasons, the presence of uncertainty. Dealing with uncertainty in robot navigation is a difficult problem because it invalidates the performance guarantees achieved in deterministic settings, while its precise effect on motion cannot be predicted. Typically, we find uncertainty embedded within the system (``process uncertainty''), in perception, and in the environment. This dissertation develops tools for dealing with process uncertainty. The developed tools lay the basis for a general framework that can be used to quantify the effect of process uncertainty on robot motion, and recover some performance guarantees for achieving motion tasks. If we could capture the variability in motion caused by process uncertainty, quantify risk, and establish performance trade-offs in its presence, we could then create consistent links between high-level objectives and low-level implementation. Such links would allow for robot navigation in real-world settings with performance certificates, a need that becomes pressing as robotics is rapidly gaining momentum in consumer applications. Dealing with uncertainty is not only important in robotics but also in general cyber-physical and biological systems; elements of this work may find applications in these domains as well. We follow a data-driven hierarchical control framework to address the need for consistent links between high-level objectives and low-level implementation. The framework has three levels: low-level control, high-level task planning, and mid-level motion generation. The benefit of the hierarchical approach is that it breaks the problem into smaller sub-parts that can be tackled more easily, and allows for available techniques at the two ends to be bridged. However, we still need to ensure compatibility and reconcile the two ends in face of uncertainty. To tackle these challenges, we focus on the mid level and propose two new important components: first, simple abstract models (``templates''), which are data-driven, ensure compatibility; second, a data-driven probabilistic framework recovers some guarantees that the policies prescribed in the high (cyber) level are implementable in the low (physical) level. As it turns out, templates must be reconciled with the physics of the system through experimental data, and this is the key to achieving consistency under uncertainty. The main ideas of the approach are fixed using a particular application area: miniature legged robots. The reduction in scale magnifies the effect of uncertainty, and thus miniature legged robots provide a suitable testbed for the proposed framework; indeed, uncertainty enters naturally (e.g., inherent uncertain leg-ground interactions), while its effect on robot motion is clearly visible. By applying the framework to this area, we enable real-time robot navigation and control at the miniature scale. The latter pushes the limits of what palm-sized crawling robots can achieve, and aids in shaping their potential in applications including building/pipe inspection, search-and-rescue, and wildlife monitoring. Overall, we provide key components of a hierarchical framework that can potentially be used for approaching more general problems of planning, navigation, and control in the presence of uncertainty. The novel components are: (i) consistent data-driven templates that ensure compatibility among the different levels of the framework; and (ii) a probabilistic framework that reconciles high-level task planners and low-level motion controllers in the presence of uncertainty. Together, these components advance the state-of-the-art in planning, navigation, and control at small scales under uncertainty, and when applied to the realm of miniature legged robots, they offer tangible benefits with regards to motion capabilities for such platforms.University of Delaware, Department of Mechanical EngineeringPh.D