Introduction to Autonomous Robotic Systems: Scope, Architecture, and Book Roadmap

The proliferation of autonomous robotic systems in the past two decades has fundamentally reshaped both industrial production and daily human life, marking a transition from specialized, pre-programmed automation to adaptive, environment-aware agents capable of operating in unstructured, dynamic settings. Unlike early industrial manipulators confined to fixed, structured production lines with no capacity for environmental adaptation, modern autonomous robots integrate suites of sensors, onboard computation, and closed-loop control pipelines to perceive their surroundings, reason about task goals, and execute actions without continuous human intervention. This chapter frames the technical scope of the present work, situates autonomous robotics within its historical and technological context, outlines the canonical system architecture that underpins all modern implementations, previews key application domains that drive current research and development, and presents the mathematical and engineering prerequisites required to engage with the subsequent technical content.

Evolution of Autonomous Robotics: From Industrial Manipulators to General-Purpose Systems

The history of autonomous robotics can be divided into four distinct eras, each defined by breakthroughs in sensing, computation, and algorithmic capability that expanded the operational envelope of robotic systems. The first era (1960–1980) was dominated by industrial manipulators, beginning with the 1961 deployment of Unimate, the first commercially produced industrial robot, at a General Motors production line in New Jersey[^1]. These early systems were non-autonomous by modern definitions: they executed pre-programmed, point-to-point motion trajectories with no feedback from the external environment, relying on hard-coded teach pendants for programming and physical guarding to ensure safety. The core technological limitation of this era was the absence of closed-loop environmental interaction: robots could not adapt to variations in part position, tool wear, or unexpected obstacles, making them suitable only for highly repetitive tasks in fully controlled environments.

The second era (1980–2000) saw the emergence of the first general-purpose mobile robots, most notably Shakey the Robot, developed by SRI International between 1966 and 1972[^1]. Shakey integrated a television camera, laser rangefinder, and bump sensors, and ran on a suite of algorithms for visual perception, path planning, and task execution, enabling it to navigate between rooms and manipulate simple objects in a controlled office environment. While Shakey’s operational speed was extremely low (it could take up to several hours to complete a simple navigation task) and its reliability was poor, it established the core architectural blueprint for all future autonomous systems: a closed-loop pipeline of perception, planning, and control. This era also saw foundational theoretical advances in probabilistic state estimation, including the development of the Kalman filter for linear systems and its nonlinear extensions, which remain core components of modern autonomous robot stacks.

The third era (2000–2020) was defined by the commercialization of mobile autonomous systems, driven by exponential improvements in sensor cost-performance, computing power, and algorithmic maturity. The 2004 and 2005 DARPA Grand Challenges catalyzed investment in autonomous driving technology, with winning teams demonstrating that LiDAR-based SLAM, probabilistic planning, and model predictive control could enable reliable navigation in unstructured off-road environments[^2]. Concurrently, the 2012 breakthrough of deep convolutional neural networks (CNNs) for image classification revolutionized robotic perception, enabling robust object detection, semantic segmentation, and 3D scene understanding from visual and LiDAR data. By 2020, level 2 (partial driving automation) systems had achieved 35% global market penetration in new passenger vehicles, and level 4 (high driving automation) autonomous vehicles were operating commercially in controlled environments including ports, mines, and urban delivery routes[^3]. This era also saw the widespread adoption of the Robot Operating System (ROS), which standardized software interfaces for robotic hardware and algorithms, accelerating system integration and prototyping.

The current, fourth era (2020–present) is characterized by the integration of foundation models—large language models (LLMs), vision foundation models, and multimodal models—into robotic autonomy stacks, enabling capabilities that were previously infeasible with modular, hand-engineered pipelines. Unlike earlier systems that relied on narrow, task-specific perception and planning modules, foundation models provide zero-shot semantic understanding, commonsense reasoning, and cross-modal generalization, enabling robots to interpret open-ended user commands, adapt to unseen environments, and perform long-horizon manipulation tasks without explicit re-programming. Concurrent advances in neural rendering, including neural radiance fields (NeRF) and 3D Gaussian splatting, are enabling high-fidelity 3D scene reconstruction and semantic querying from sparse sensor data, addressing long-standing limitations of traditional geometry-based SLAM in low-texture or dynamic environments. As of 2024, the global autonomous robot market is valued at $48 billion, with a projected compound annual growth rate (CAGR) of 21.2% through 2030, driven by demand for warehouse automation, last-mile delivery, and collaborative industrial robots[^4].

<tag>历史演进</tag>

Core Scope of This Book: The Three Pillars of Perception, Planning, and Control

Autonomous robotics is an inherently interdisciplinary field, spanning mechanical engineering, computer science, electrical engineering, and applied mathematics. To provide a focused, in-depth treatment of the core technical content, this book concentrates on the three interdependent pillars that form the functional core of any autonomous robotic system: perception, planning, and control. We explicitly exclude coverage of low-level mechanical design, embedded hardware development, and motor driver implementation, as these topics are well-covered in dedicated robotics textbooks and are not core to the algorithmic and mathematical foundations emphasized here.

Perception

Perception refers to the set of algorithms and models that enable a robot to infer its own state (ego-state) and the state of its surrounding environment from raw sensor data. Core perception capabilities covered in this book include:

• **State estimation**: Fusing data from heterogeneous sensors (IMU, wheel encoders, LiDAR, cameras, GPS) to produce accurate, low-latency estimates of the robot’s position, velocity, orientation, and angular velocity.
• **Scene understanding**: Extracting semantic information from sensor data, including 2D/3D object detection, semantic segmentation, instance segmentation, and depth estimation, to identify obstacles, navigable space, and task-relevant objects.
• **Mapping**: Constructing spatial representations of the environment, ranging from sparse feature maps for localization to dense semantic maps for navigation and manipulation.

Planning

Planning refers to the set of algorithms that generate a sequence of actions or a feasible trajectory for the robot to achieve its high-level task goals, subject to dynamic and kinematic constraints. We distinguish three tiers of planning covered in this book:

• **Task planning**: Decomposing high-level user goals (e.g., “deliver a package to the 3rd floor conference room”) into a sequence of atomic subtasks (e.g., “navigate to elevator”, “call elevator”, “enter elevator”, “navigate to conference room”), accounting for resource constraints and dynamic environment changes.
• **Behavioral decision-making**: Selecting high-level behavioral modes (e.g., “follow lane”, “change lane”, “yield