Recent progress in robotic systems has significantly advanced robot functional capabilities, including perception, planning, and control. As robots are gaining wider applications in our society, they have started entering our workplace and interacting with us. This leads to new challenges for robots: they are expected to not only be more functionally capable automatic machines, but also become human-compatible, which requires robots to make themselves competent agents to work for people and collaborative partners to work with people on diverse tasks. The capability to planning under uncertainty lies at the core to achieving this goal. The aim of this dissertation is to develop new approaches that improve the autonomy and intelligence of robots to enable them to reliably work for and with people. Especially, this dissertation investigates uncertainty reduction and the planning under various types of uncertainty with the focus on three related topics, including distributed filtering, informative path planning, and planning for human-robot interaction.
In the first topic, the dissertation studies uncertainty reduction via distributed filtering using networked robots. We consider the distributed version of the generic Bayesian filter. Two new methods of measurement exchange among networked robots are proposed, which enable the dissemination of robots’ sensor measurements in time-invariant and time-variant communication networks. By using such methods, the communication burden of the robot network can be significantly reduced compared to traditionally used methods. Based on these measurement exchange methods, we develop two distributed Bayesian filters for time-invariant and time-variant networks. It has been proved that the proposed distributed Bayesian filter can achieve consistent estimation. The application in target localization and tracking is presented.
In the second part, the dissertation focuses on planning under the uncertainty of target position and motion model. This part investigates the path planning of a mobile robot to autonomously search and localize/track a static/moving target. We first study the case of linear Gaussian sensing and mobility models. A path planning approach based on model predictive control (MPC) is proposed, which uses a modified Kalman filter for uncertainty prediction and a sequential planning strategy for path generation. We then investigate the path planning in a dynamic environment, with the sensor using a binary model. A closed-form objective function for the MPC-based path planner is proposed, which significantly reduces the computational complexity. The safety of robot is enforced by using a barrier function in the objective function of MPC.
The first two topics concentrate on making robots autonomously work for people. In the third topic, the dissertation addresses the demands to make robots work with people and achieve coordination. We first consider the planning of robots under the uncertainty of humans’ trajectory in a human-following application, where the robot needs to generate a path to follow a person in a safe and comfortable way. We propose a model-based human motion prediction approach using the principle of interacting multiple model estimation. A path planner based on nonlinear MPC is then developed for the robot to generate human-following paths, which takes into account the safety and comfort of the accompanied person.
We then investigate the planning of robots under the uncertainty of humans’ internal states, including their intention and belief. Especially, the task planning in the human-robot collaboration is considered. We develop an adaptive task planning scheme that allows a robot to use motion-level inference to understand a human partner’s plan and then adjust its task-level plan to coordinate with the person. In addition, we model a person’s inference process and develop a task planning approach for a robot to generate human-predictable plans, which aims to reduce the misalignment between people’s belief and robots’ plan.