Introduction to the Robotics Perception-Planning-Control Paradigm

Modern robotics systems have evolved from pre-programmed, structured-environment industrial manipulators to autonomous agents capable of operating in unstructured, dynamic, and partially observable environments, a transition fundamentally enabled by the maturation of the perception-planning-control (PPC) paradigm. This closed-loop operational framework, which integrates environmental sensing, decision-making, and physical actuation, has become the de facto standard architecture for nearly all advanced robotic platforms, from self-driving vehicles to humanoid service robots and dexterous industrial manipulators. Unlike early robotic systems that relied on hard-coded trajectories and limited sensory feedback, contemporary PPC frameworks enable robots to adapt to unforeseen environmental changes, recover from actuation errors, and execute complex long-horizon tasks with minimal human intervention. The core value of the PPC paradigm lies in its modular yet tightly coupled design: each module addresses a distinct functional requirement while exchanging information with other modules to form a robust closed loop that accounts for environmental uncertainty, actuation latency, and model mismatch. This chapter first traces the historical evolution of the PPC paradigm, then formally defines the functional boundaries of each module, analyzes their closed-loop interaction mechanisms, outlines prerequisite knowledge for intermediate readers, and illustrates implementation nuances across three representative application scenarios, before introducing the structure of this book.

1.1 Historical Evolution of the Robotics Operational Paradigm

The conceptual foundation of the modern PPC paradigm can be traced back to the 1960s, when SRI International developed Shakey the Robot, the first general-purpose mobile robot capable of reasoning about its actions[^1]. Shakey’s architecture was a direct precursor to the modern PPC triad: it used a ranging laser and monocular camera for environmental perception, a STRIPS (Stanford Research Institute Problem Solver) planner for generating action sequences to achieve high-level goals (e.g., "move the box to the room"), and a low-level controller to drive its wheeled locomotion and manipulator. While Shakey could only operate in highly simplified, controlled indoor environments due to limited sensing and computing power, its three-layer architecture established the core logic of separating environmental understanding, decision-making, and actuation that persists today.

The 1980s saw the widespread adoption of industrial manipulators in manufacturing, but these early systems relied almost entirely on pre-programmed, open-loop trajectories with minimal sensory feedback. The dominant operational paradigm was "teach-and-playback," where operators manually guided the robot through desired poses, and the robot repeated the trajectory without environmental awareness. This changed in the 1990s with the introduction of force/torque sensors and visual servoing, which added basic perception and closed-loop control capabilities to industrial robots, enabling them to adapt to small variations in workpiece placement. However, these systems still operated in fully structured environments with no need for high-level planning or semantic perception.

The early 2000s marked a turning point for the PPC paradigm, driven by the DARPA Grand Challenge series, which tasked teams with building autonomous ground vehicles capable of navigating desert and urban routes without human intervention. The 2005 Grand Challenge was won by Stanford’s Stanley, which combined GPS-based global localization, 2D laser SLAM for local mapping and obstacle detection, A* global path planning, and PID-based longitudinal and lateral control to complete the 212km desert course in 6 hours and 54 minutes, with an average speed of 19km/h[^3]. The 2007 Urban Challenge required vehicles to navigate city streets, obey traffic rules, and interact with other human-driven vehicles, driving the adoption of multi-sensor fusion perception, dynamic obstacle trajectory prediction, and model predictive control (MPC) for robust trajectory tracking. These competitions demonstrated that the PPC paradigm could scale to complex, dynamic real-world environments, catalyzing research and industrial adoption across all robotics domains.

The past decade has seen the PPC paradigm evolve further with the integration of deep learning, enabling robots to perform semantic perception, end-to-end policy learning, and sim-to-real transfer. While traditional PPC systems relied on hand-engineered features and explicit geometric models, learning-augmented PPC frameworks can generalize to unseen environments and tasks by leveraging large-scale datasets. For example, modern autonomous driving systems use convolutional neural networks (CNNs) for 3D object detection and semantic segmentation, transformer-based models for dynamic obstacle trajectory prediction, and reinforcement learning (RL) for behavior planning, while retaining traditional model-based control modules for safety-critical trajectory tracking. This hybrid approach balances the generalizability of learning-based methods with the safety and interpretability of model-based PPC modules, and is now the dominant architecture for state-of-the-art robotic systems.

1.2 Core Definitions and Functional Boundaries of the PPC Triad

While the PPC paradigm is widely adopted, clear functional boundaries between the three modules are critical for modular development, debugging, and certification, especially for safety-critical applications such as autonomous driving and industrial automation. Each module has distinct input/output interfaces and core functional requirements, as formalized below.

1.2.1 Perception Module

The perception module is responsible for extracting structured, actionable information about the robot’s internal state and external environment from raw, noisy sensor data. Its core functional scope includes: 1. State estimation: Fusing multi-sensor data to estimate the robot’s pose, velocity, and internal states (e.g., joint angles for manipulators, battery level for mobile robots). The most common formulation is Bayesian filtering, where the posterior state distribution $p(x_k | z_{1:k}, u_{1:k})$ is recursively updated using the motion model $p(x_k | x_{k-1}, u_{k-1})$ and observation model $p(z_k | x_k)$, where $x_k$ is the robot state at timestep $k$, $z_{1:k}$ are historical sensor observations, and $u_{1:k}$ are historical control inputs. For linear Gaussian systems, this reduces to the Kalman filter update rule: $$\hat{x}_{

$$\hat{x}_{k|k} = \hat{x}_{k|k-1} + K_k (z_k - H_k \hat{x}_{k|k-1})$$ 其中$K_k$为卡尔曼增益，$H_k$为观测矩阵，$\hat{x}_{k|k-1}$为基于运动模型的一步预测状态。除状态估计外，感知模块还承担环境建图与语义理解任务：同步定位与建图（SLAM）通过联合优化机器人位姿序列与地图参数，解决“定位-建图”的耦合问题，其因子图优化目标可形式化为： $$\min_{x_{1:T}, m} \sum_{t=1}^T \left( \| x_t - f(x_{t-1}, u_t) \|_{\Sigma_u}^2 + \sum_{z \in Z_t} \| h(m_{l_z}, x_t) - z \|_{\Sigma_z}^2 \right)$$ 其中$x_t$为$t$时刻机器人位姿，$m$为环境地图，$f$为运动学模型，$h$为观测模型，$\Sigma_u$与$\Sigma_z$分别为运动噪声与观测噪声的协方差矩阵。当前主流开源方案ORB-SLAM3即基于该优化框架实现，支持视觉、激光雷达、IMU多传感器融合，在室内场景下的定位精度可达厘米级。此外，基于深度学习的3D目标检测、语义分割等任务已成为自动驾驶、服务机器人等场景的感知标配，例如BEV（鸟瞰图）感知范式通过将多视角摄像头图像转换为统一的鸟瞰图特征，实现了3D目标检测mAP超过90%的精度，远高于传统手工特征方法的65%平均水平。

感知模块的核心挑战来自环境不确定性与传感器噪声：激光雷达在雨量10mm/h的场景下点云密度会下降40%，有效探测距离缩短30%；摄像头在逆光、夜间场景下的检测准确率会下降20%以上；动态障碍物（如横穿马路的行人、移动的工业AGV）的观测存在运动模糊，导致位姿估计偏差。为应对这些挑战，现代感知模块普遍采用多传感器融合架构，通过冗余观测降低单传感器失效的风险，例如自动驾驶车辆通常同时搭载激光雷达、摄像头、毫米波雷达、IMU四类传感器，利用激光雷达的精确几何测量、摄像头的语义识别能力、毫米波雷达的恶劣天气穿透能力、IMU的高频姿态更新能力，实现全场景下的稳定感知。

1.2.2 Planning Module

规划模块接收感知模块输出的机器人状态估计、环境地图与障碍物信息，结合高层任务目标，生成满足安全约束、动力学约束、任务约束的可执行动作序列或轨迹。根据决策时间尺度与抽象层级，规划模块可划分为三层子架构：行为规划层（Behavior Planning）、运动规划层（Motion Planning）、轨迹优化层（Trajectory Optimization），各层的功能边界与输入输出对比如表1.1所示。

表1.1 规划模块三层子架构的功能对比 | 子模块 | 时间尺度 | 抽象层级 | 核心输入 | 核心输出 | 典型算法 | |----------------|----------|---------