Search

This thesis deals with sampled-data implementations of continuous-time, non-linear control systems. The basis for the analysis is a static, continuous-time feedback law for non-linear, affine systems with bounded input gain. The sampled-data implementation is obtained from the discretization of the control via a sample-and-hold-process. With the incorporation of the aspect of robustness, a theoretical framework is created which supersedes previous work concentrating on stability. Bounding constraints for the closed-loop differential system allow uncertainty and disturbances to be considered. Other assumptions for the continuous-time control are Lipschitz continuity, exponential decay outside a compact set and existence of a Lyapunov function. The important parameter for the discretization analysis is the sampling time; fast sampling implies robust stability. The controller sampling residual, the difference between the discretized and the original control, is of key interest within a Lyapunov-type stability analysis; suitable norms, such as the Euler norm, are chosen to find upper bounds for the sampling residual. The generalization of a result from linear to non-linear sampled-data control permits the application of the Lp-norm. The theoretical framework is also suitable for dynamic control systems and the investigation of computational delays. The analysis approaches are demonstrated for two different robust control. strategies .based on sliding-mode approaches. A state-feedback sliding-mode-based control extends ideas for smoothing discontinuous sliding-mode control components by introducing a cone-shaped sliding-mode layer. A non-smooth Lyapunov function is used to prove stability of the discretized control. An observer-based tracking control improves a previous control scheme by considering a class of non-minimum phase and relative-degree-zero plants. Simulation and numerical fast-sampling analysis results are provided for all developed discretization and sliding-mode-based control techniques in application to non-trivial examples. The simulation of a highly non-linear, large-scale chemical plant for benzene production with non-minimum phase and relative-degree-zero characteristics proves the effectiveness of sliding-mode output control.

With the development of industrialization, the working environment of robotics gradually becomes complex, diverse, and fast. Most manipulators at present are still designed for simple action repetition, which means that the working environment is determined and the target should be relatively fixed. Therefore, they lack the ability to perceive the surrounding environment. The main purpose of this thesis is to develop consensus-based training and deep reinforcement learning methods that enable robot arms to interact with the environment autonomously. First of all, a model-free off-policy actor-critic based deep reinforcement learning method is proposed to solve the classical path planning problem of a UR5 robot arm. The proposed method not only guarantees that the joint angle of the UR5 robotic arm lies within the allowable range each time when it reaches the random target point, but also ensures that the joint angle of the UR5 robotic arm is always within the allowable range during the entire episode of training. Moreover, a self-supervised vision-based deep reinforcement learning method that allows robots to pick and place objects effectively and efficiently when directly transferring a training model from simulation to the real world is demonstrated. A height-sensitive action policy is specially designed for the proposed method to deal with crowded and stacked objects in challenging environments. The training model with the proposed approach can be applied directly to a real suction task without any fine-tuning from the real world while maintaining a high suction success rate. It is also validated that the training model can be deployed to suction novel objects in a real experiment with a suction success rate of 90% without any real-world fine-tuning. Additionally, an algorithm that combines actor-critic based off-policy method with consensus-based distributed training is proposed to deal with multi-agent deep reinforcement learning problems. Specially, a convergence analysis of a consensus algorithm for a type of nonlinear systems with a Lyapunov method is developed, and this result is used to analyse the convergence properties of the actor and the critic training parameters. To validate the implementation of the proposed algorithm, a multi-agent training framework is proposed to train each UR5 robot arm to reach the random target position. Experiments are provided to demonstrate the effectiveness and feasibility of the proposed algorithm. Finally, a Consensus-based Sim-and-Real deep reinforcement learning algorithm is developed for manipulator pick-and-place tasks. Agents are trained in both simulators and the real environment simultaneously to get the optimal policies for both sim-and-real worlds. The proposed algorithm saves required training time and shows comparable performance in both sim-and-real worlds.

The advent of complex machines such as robots and autonomous vehicles has opened the possibility of close interaction and cooperation with human users, for which safety becomes crucially important. By considering autonomous vehicles as robots (i.e. autonomous agents), a bridging of both areas can be done by bringing methods from human-robot interaction to autonomous vehicular agents' interaction. Drawing ideas from Human Robot Interaction (HRI), the understanding and estimation of human behaviour can ensure that a task is being performed by a human the way a robot/vehicle is expecting. Our results provide useful estimations and task-related predictions of human behaviour that can be used in automatic control loops to ensure safe cooperative actions. Different modelling objectives were explored to gain insight into human intention and readiness in terms of task performance; task modelling at a temporal manoeuvre-level, task performance whilst multitasking and cognitive load (i.e. mental workload) were studied. Physiological signals from different sources such as skeletal tracking sensors, heart monitors and eye trackers were used to acquire knowledge about the user's intentions and behaviours (e.g. cognitive load). Data-driven methods, from shallow to deep learning ones, were used to relate data from multi-modal heterogeneous sources (e.g. human driver, vehicle) to expected behaviour, in order to create meaningful relations to task intention and cognition. Most relevant results included manoeuvre identification and prediction for known and unknown test subjects using skeletal tracking data. Outcomes showed relations between driving behaviour and mental workload using car dynamics, heart rate and eye gaze metrics, predicting driving performance up to 1 second ahead of the actual actions.

In complex networked systems, the structure of the underlying communication between individual agents can have profound impacts on the performance and dynamics of the system as a whole. For example, processes such as consensus can occur at faster or slower rates depending on the structure of the communication graph, and the synchronisation of coupled chaotic oscillators may be unachievable in one configuration, when it is achievable in another. As such, it is vital for agents within a complex networked system to be able to make estimates of the properties of the network as a whole, and be able to direct their own actions to modify these properties in a desirable way, even when they are only able to communicate with their direct neighbours, and have no global knowledge of the structure of the network. In this thesis we explore decentralised strategies by which individual agents in a network can make estimates of several functions of the graph Laplacian matrix eigenvalues, and control or optimise these functions in a desired manner, subject to constraints. We focus on the following spectral functions of the graph Laplacian matrix, which determine or bound many interesting properties of graphs and dynamics on networks: the algebraic connectivity (the smallest non-zero eigenvalue), the spectral radius (the largest eigenvalue), the ratio between these extremal eigenvalues (also known as the synchronisability ratio), the total effective graph resistance (proportional to the sum of the reciprocals of non-zero eigenvalues), and the reduced determinant (the product of the non-zero eigenvalues).