Exploring Active Inference and Model Predictive Path Integral Control: A Journey from Low-Level Commands to Task and Motion Planning

In an ever-evolving society, the demand for autonomous robots equipped with human-level capabilities is becoming increasingly imperative. Various factors, such as an aging population and a shortage of labor for repetitive and physically demanding tasks, have underscored the need for capable autonomous robots to assist us in our daily activities. However, despite the recent advancements in robotics, the field still faces significant challenges in delivering on its promises of developing general-purpose robots with human-level capabilities for everyday tasks. This thesis aims to develop control algorithms at different levels of abstraction to achieve more robust, adaptive, and reactive robot behavior for long-term tasks in dynamic environments.

Since our ultimate goal is to achieve human-level performance, a natural starting point is to investigate theories of human intelligence and how they can be applied to real robots, such as mobile manipulators. In this regard, one prominent theory is Active Inference, a popular and influential concept that can explain a wide range of cognitive functions, from motor control to high-level decision-making. Active Inference was developed based on the free-energy principle providing an explanation for embodied perception-action loops. While the free-energy principle and Active Inference have garnered significant attention among neuroscientists, their application to robotics remains largely unexplored, presenting an exciting avenue for research in this thesis. At the same time, it is also important to recognize that we should not confine ourselves solely to theories of human intelligence and their inherent limitations. Machines and humans are built upon fundamentally different structures, which opens up possibilities for alternative approaches. Consequently, this thesis also investigates the use of Model Predictive Path Integral Control (MPPI), which stems from a different formulation of free-energy that is not bound to biological assumptions. By exploring the application of Active Inference to low-level robot control and task planning, as well as the utilization of MPPI for motion planning, this thesis provides advancements in robot control at different levels of abstraction. More concretely, this thesis contributes to the following four areas: 1) Lowlevel adaptive and fault-tolerant control, 2) Reactive high-level decision making, 3) Contact-rich motion planning, and 4) Reactive task and motion planning (TAMP)…