Titolo della tesi: Optimization-Based Methods for Real-Time Generation of Safe Motions in Mobile Robots
Having robots operating in unstructured and dynamically changing environments is a challenging task that requires advanced motion generation approaches that are able to perform in real-time while maintaining the robot and environment safety.
The progress in the field of numerical optimization, as well as the development of tailored algorithms, made Nonlinear Model Predictive Control (NMPC) an appealing candidate for real-time motion generation. By considering the robot model as prediction model and through appropriate constraints on the robot states and control inputs, NMPC can enforce safety to the resulting motion in a straightforward way.
This thesis addresses the problem of real-time generation of safe motions for mobile robots and mobile manipulators. The different structure of the considered robots introduces different safety risks during the robot motion and so the motion generation problem for each robot is addressed in separate parts of this thesis.
The first part addresses the problem of motion generation for mobile robots navigating in environments populated by static and/or moving obstacles. For the generation of the desired motion, real-time NMPC is used. We argue that, in order to tackle the risk of collision with the environment, traditional distance-based approaches are incapable of maintaining the robot safety when the NMPC uses relatively short prediction horizons. Instead, we propose two NMPC approaches that employ two alternative collision avoidance constraints. The first proposed NMPC approach is applied for a scenario of safe robot navigation in a human crowd. The NMPC serves as motion generation module in a safe motion generation framework, complete with a crowd prediction module. The considered collision avoidance constraint is built upon an appropriate Control Barrier Function (CBF). The second NMPC approach is applied to a scenario of robot navigation among moving obstacles, where the dynamics of the considered robot are significant. The proposed collision avoidance constraint is built upon the notion of avoidable collision state, which considers not only the robot-obstacle distance but also their velocity as well as the robot actuation capabilities. The simulation results indicate that both methods are effective and able to maintain the robot safety even in cases where their purely distance-based counterparts fail.
The second part of the thesis addresses the problem of safe motion generation for mobile manipulators, call to execute tasks that may require aggressive motions. Here, in addition to the risk of collision with its environment, the robot, consisting of multiple articulated bodies, is also susceptible to self-collisions. Moreover, fast motions can always result to loss of balance. To solve the problem, we propose a real-time NMPC scheme that uses the robot full dynamics, in order to enforce kinodynamic feasibility, while it also considers appropriate collision and self-collision avoidance constraints. To maintain the robot balance we enforce a constraint that restricts the feasible set of robot motions to those generating non-negative moments around the edges of the support polygon. This balance constraint, inherently nonlinear, is linearized using the NMPC solution of the previous iteration. In this way, we facilitate the solution of the NMPC in real-time, without compromising the robot safety.
Although the proposed NMPC is effective when applied to MM with low degrees of freedom, when the robot becomes more complex the use of its full dynamic model as prediction model in an NMPC can lead unacceptably large computational times that are not compatible with the real-time requirement. However, the use of a simplified model of the robot in an NMPC can compromise the robot safety. For this reason, we propose an optimization-based controller equipped with balance constraints as well as CBF-based collision avoidance constraints. The proposed controller can serve as an intermediate between a motion generation module that does not consider the robot full dynamics and the robot itself in order to ensure that the resulting motion will be at least safe. Simulation results indicate the effectiveness of the proposed method.