Planning, Acting and Monitoring in Robotics Environments

As hardware and software costs continue to drop, consumer robotics has become easier to develop and increasingly attractive to industry. As a result, autonomous mobile robots are being deployed in unstructured public environments such as malls, hotels, and museums. In these settings, robots face unpredictable changes and failure modes and when they encounter situations outside their design assumptions, they may fail and require human intervention. Mitigating failures in the field typically requires updating the software to address newly observed issues and re-testing previously working scenarios. This process is time-consuming and difficult to carry out exhaustively in practice, especially as users demand more human-like robots whose behavior must be updated at runtime or extended with new capabilities. A key challenge is therefore ensuring that the robot’s executed behavior matches its specified behavior, particularly for complex, large-scale systems. In this thesis I addresses these challenges by presenting (i) a model-based framework that generates executable code from deliberation models, enabling off-line correctness verification and on-line execution monitoring and (ii) a task-level planner for operation in crowded environments. Together with my colleagues at IIT, I developed a model-based code generator that generates executable code from an XML-based language based on the interpretation of standard SCXML code. My contributions also include implementing the validation use cases, defining their verification and monitoring properties, and grounding MOON in concrete robotic scenarios. The robotics domain expertise gained from the robotic tour-guide use case guided the development of the monitoring approach, while feedback from the validation activities informed design decisions for the shared tools. To address behavioral correctness, deliberation components are modeled at a conceptual level using Model-Based Development (MBD). Executable code is therefore automatically generated from these models, allowing the resulting behavior to be verified with existing model checkers and monitored at runtime using a monitoring tool. For deployment in congested areas, I developed a probabilistic tour planner that leverages learned Circular Linear Flow Field (CLiFF) maps to predict pedestrian motion. The planner formulates and solves a Markov Decision Process (MDP) on-line to route the robot while adapting to newly observed pedestrians. We then integrated the generator and the software mentioned above into a robotic tour guide developed within the CONVINCE consortium and evaluated in simulation and real-world experiments.