Chapter 1
Introduction to Design Automation and Intelligent Robotic Systems

1.1 Background of Design Automation

1.1.1 Historical Evolution

Design automation is an important branch of knowledge automation and is listed by McKinsey as one of the second largest and most disruptive technologies that will determine the future economy. It plays a crucial role in the design of intelligent robots, especially in the design of the robot's body, control system, and vision system. By utilizing technologies such as genetic programming (GP), evolutionary computing, deep learning, reinforcement learning, and causal inference, design automation can significantly improve design efficiency, avoid the repetitive trial-and-error process in traditional design methods, and thus promote the automatic discovery of innovative designs.

For example, famous robot companies such as ABB attach great importance to research in the field of modular design automation (MODENA) for intelligent robots. ABB's robot research and development team has established a long-term cooperative relationship with Professor Peter Krus's team at Linköping University in Sweden and is committed to in-depth research on the automatic design optimization problem of robot systems. In their publicly published papers, they clearly define the concept of "Design Automation" [1]. In their cooperative research, design automation of geometric graph models is at the core, and the classic multi-objective evolutionary algorithm (MOEA) is mainly used to optimize the design of robot systems [1]. However, since the classical MOEA is relatively simple in the constraint-handling mechanism, and it is difficult to effectively solve large-scale and complex constrained multi-objective optimization problems (CMOPs), systems designed using this method are often limited in scale expansion.

Another important representative work in MODENA for intelligent robots comes from Professor Hod Lipson of Columbia University in the United States. They used evolutionary computing methods to automatically design a movable robot system on a computer and physically realized it through 3D printing technology, creating what is claimed to be the world's first robot system automatically designed and manufactured by a computer [2]. Professor Hod Lipson's team has continued to make many new progress in this field [3, 4].

The group led by Professor Erik Goodman, the founding director of the BEACON National Center for Scientific and Technological Innovation at Michigan State University, has also carried out pioneering research in the field of electromechanical system design automation. They mainly adopted a method combining GP and bond graph and genetic programming (BGGP) and successfully achieved the automatic design of a class of electromechanical systems [5]. Subsequently, the team of Professor Clarence D. Silva, a famous authority in the field of mechatronics at the University of British Columbia in Canada, completed another doctoral dissertation and extended this research from the field of linear systems to the field of nonlinear systems. At the same time, they also proposed an important concept of mechatronic design quotient [6]. This concept can integrate multiple design goals in engineering design into one goal, so that a single-objective optimization algorithm can be used for automatic design. Subsequently, Fan et al. [7] extended the automatic design of electromechanical systems from continuous systems to hybrid electromechanical systems containing continuous dynamic systems and discrete events, and realized the parallel design of controllers and controlled objects, further promoting research in this field. In conclusion, research in the field of electromechanical system design automation has formed a certain inheritance internationally [5-8]. As a typical representative of electromechanical systems, research on MODENA for intelligent robots [9-12] combined with electromechanical system design automation [5-8] is forming a broader research direction.

For example, in the aspect of robot body design, Fan et al. [13] used the algorithm based on the push-pull search (PPS) framework [14] to conduct design optimization on a six-degree-of-freedom teaching robotic arm and obtained a design scheme better than that designed by human engineers. In the aspect of robot vision system design, Fan et al. embedded the proposed lightweight codec network model [15] into a composite mobile robot and developed a road crack detection robot for open environments. In the aspect of automatic design of robot body and controller, the team of Professor Li Fei-Fei from Stanford University proposed a new computing framework - deep evolutionary reinforcement learning [16]. Based on this framework, embodied agents can perform multiple tasks in multiple complex environments. In addition, this research also demonstrated for the first time through morphological learning the Darwin-Baldwin effect in evolutionary biology. In the aspect of automatic morphological design of swarm robots, Fan et al. [17] proposed a design automation framework for automatic generation of swarm robot morphologies. The gene regulatory network (GRN) generated by this framework not only has a simpler structure but also has better performance than the model designed by human experts. In addition, the GRN model generated by this framework can generate irregular swarm patterns in dynamic environments and thus flexibly surround and capture targets.

Although many achievements have been made in the MODENA for intelligent robots, a series of challenges still remain. For example, how to effectively integrate and optimize the integrated design of the robot's "morphology (body) - brain (controller) - eyes (vision)," and how to deal with the complex coupling relationships between various modules in multi-energy domain complex systems. Future research on design automation needs to make further breakthroughs in the following aspects [18]: (1) multi-angle unified modeling of intelligent robots; (2) expensive constrained multi-objective optimization method based on surrogate models; (3) robot design automation based on Digital Twins; (4) domain adaptation and generalization in design automation; (5) design decision control method integrating large models and domain agents; and (6) design of embodied intelligent robot systems based on MODENA.

1.1.2 Importance in Modern Engineering

The design of intelligent robots is inherently complex, and traditional methods often rely on tedious trial-and-error processes. This approach results in time inefficiencies and hampers innovation. Design automation effectively addresses these challenges. It represents a significant branch of knowledge automation, or the automation of knowledge work. Unlike traditional automation, which focuses on enhancing human physical capabilities, knowledge automation leverages mental activities such as cognition, reasoning, and imagination. According to a McKinsey report, knowledge automation ranks as the second most disruptive technology. The advent of technologies like ChatGPT and other generative artificial intelligence (AI) tools marks the beginning of the knowledge automation era, where tasks such as essay writing and code programming are increasingly automated. Design automation plays a crucial role in knowledge automation because it encompasses a major part of human mental and innovative activities. Consequently, design automation has become an indispensable part of modern engineering.

1.2 Intelligent Robotic Systems

1.2.1 Definition and Components

Intelligent robotic systems are the specific application of intelligent machines in the field of robotics. They are complex, computer-controlled robotic systems equipped with multiple visual and non-visual sensors, possessing the ability to make decisions and solve problems within their operational domain. Their design and operation are based on the theory of hierarchical intelligent control systems and general systems theory, utilizing both analytical and heuristic methods for system modeling and control.

The components of intelligent robotic systems are divided into three hierarchical levels. The organizational level is basically structured after a knowledge-based system [19]. It is primarily responsible for functions such as reasoning, planning, decision-making, feedback, and long-term memory exchange. At this level, information processing resembles certain functions of the human brain, such as the processing and application of knowledge. For instance, when planning a task, it starts with high-level information, considering the goals of the task, available resources, and various constraints to generate a series of possible action plans. The second level is the coordination level, which serves as a crucial interface between the organizational and execution levels. It processes real-time information and generates specific sub-tasks based on the plans developed by the organizational level. This layer consists of multiple coordinators, each with a fixed structure and specific function. For example, the visual system coordinator is responsible for processing visual information, and the sensor system coordinator handles information received from various sensors. Communication between these coordinators is managed by a scheduler, whose structure adapts based on instructions from the organizational level. The final level is the execution level, responsible for carrying out specific control...