Neural network and metaheuristic based learning and control of articulated robotic agents

Abstract

Articulated robotic agents have become an integral part of industrial automation, and rapidly finding their way into the concept of futuristic, intelligent spaces, i.e., smart homes, assistive rehabilitation, and healthcare. As such, they have been a topic of great interest for researchers from the field of the control system, mechanics, electronics, and system design. An articulated robot consists of two main components; a mechanical system consisting of joints and an end-effector to manipulate objects. This thesis explores different aspects of the design and control system of these two components. For example, this study analyzed the prospects of attaching a flexible gripper as an end-effector of the articulated robots. Such a flexible gripper has the advantage of inherent safety as compared to their rigid counterparts. Similarly, several techniques to accurately and robustly control the joint-motion of the articulated robots, while complying with several real-world constraints, are also comprehensively presented in this thesis. First, we present a novel nature-inspired metaheuristic optimization framework to numerically solve the nonlinear optimization problem in a numerically efficient manner. This optimization framework is of paramount importance to this thesis because it is later applied to formulate efficient algorithms to control the motion of joints and end-effector. The metaheuristic algorithm is inspired by the natural behavior of beetles and mathematically model their food foraging behavior. The optimization algorithm is converted into an equivalent neural network architecture to take advantage of hardware and software optimizations provided by modern neurocomputing systems. Second, we applied the formulated optimization algorithm to model-free motion control of articulated robots. The articulated robots employed in the industry usually have redundant joints, i.e., more joints than the dimension of task-space. The redundancy can be utilized to achieve other design goals, e.g., power efficiency, torque minimization, obstacle avoidance, etc. However, the inverse kinematic model for such articulated robots cannot be expressed analytically, and numerical techniques are required to control its joint-motion. We proposed a model-free control technique to control the joint motion without relying on the inverse kinematic model or Jacobian matrix of the robot. Third, we then applied the formulated optimization algorithm to motion control of articulated robots while avoiding the collision with obstacles present in the surrounding. We proposed an optimization-centric approach that unifies the task of obstacle avoidance and trajectory tracking into a single optimization problem using the penalty-term method. The penalty term forces the optimizer to actively avoid the obstacles by adding a penalty value into the objective function when the articulated robots approach an obstacle. An important feature of the proposed algorithm is that it does consider the obstacle to a point object; instead, it considers their 3D geometry and uses GJK (Gilbert-Johnson-Keerthi) algorithm to estimate the distance between the obstacle and articulated robot. Fourth, we considered the problem of the remote center of motion (RCM) constraints of articulated agents and their link to surgical robots. Similar to the case of obstacle avoidance, we formulated an optimization problem incorporating the kinematic control and RCM constraints into a single objective function using the penalty term approach. The metaheuristic optimization algorithm discussed earlier is then applied to solve the optimization problem. The performance of the proposed controllers is analyzed using the simulated model of the IIWA-14 articulated robots, and comprehensive results are presented to demonstrate the efficacy. Fifth, we present an experimental study on model-free control of soft robots, i.e., robots made from a flexible material. These robots have the potential of being used as an end-effector of an articulated agent. Due to their flexible structure and compliance, they have the inherent advantage of safety in the handling of fragile objects in an industrial environment. A soft gripper can easily handle delicate objects, e.g., fruits, and toys, without requiring extremely complex sensing and actuation mechanisms. However, the softness and flexibility also make them extremely challenging to mathematically model and regulate their dynamic behavior. We present an experimental study to analyze the effect of model-free controllers on the motion of soft robots and proposed strategies to tune the controller parameter to obtain a desirable dynamic response.