Introduction

A robot is not merely a machine that moves. In careful engineering language, a robot is an actuated, programmable mechanism that can perform intended tasks with some degree of autonomy; the international robotics vocabulary standard frames the concept around actuation, programmability, motion in an environment, and autonomy rather than around human-like appearance (ISO 8373:2021). This matters because a surgical manipulator, a warehouse mobile robot, an underwater vehicle, a legged inspection platform, and a collaborative industrial arm may look very different, yet they all raise the same central engineering question:

How can a physical system sense, decide, and act reliably in a changing world?

This book is about that question at graduate level. It treats robotics as an engineering discipline built from mechanics, mathematics, control theory, estimation, computation, perception, learning, and experimental validation. Modern robotics texts and handbooks emphasize exactly this integration: the field cannot be reduced to only mechanical design, only artificial intelligence, or only control; useful robots require these parts to work together under physical constraints (Siciliano and Khatib, 2016; Lynch and Park, 2017).

The first orientation: a robot closes a loop with the world

The simplest useful mental model of a robot is a closed-loop system. A closed-loop system measures something, compares that information with a desired behavior, and acts in a way that depends on the measurement. The term “closed” means that action affects the world, the changed world affects future measurements, and those measurements affect later actions.

For example, consider a mobile robot driving down a corridor. It uses sensors to estimate where it is, computes a steering command, moves, receives new sensor data, and updates its estimate. The robot is not simply executing a fixed list of motions. It is continuously using feedback.

A minimal robotics loop contains five ideas:

1. The body: the physical mechanism, including links, joints, wheels, propellers, legs, grippers, motors, transmissions, batteries, and structural materials.
2. The state: a mathematical description of the relevant condition of the robot and environment.
3. Sensing and estimation: the process of measuring the world and inferring quantities that cannot be measured perfectly.
4. Planning and control: the process of choosing desired motions or actions and producing commands that drive the system toward them.
5. Interaction: contact, force exchange, uncertainty, disturbance, and communication with humans, objects, terrain, fluids, or other robots.

A state is not the same thing as the entire physical universe. It is a chosen mathematical summary. For a two-link planar robot arm, the state might include the two joint angles and two joint velocities. For an autonomous car, the state might include position, heading, speed, acceleration, and nearby obstacle estimates. For a quadrotor, it may include position, velocity, orientation, angular velocity, and sensor biases. Choosing a good state representation is one of the quiet skills of robotic engineering.

This book repeatedly returns to the same pattern: define the system, define the state, define the task, define uncertainty and constraints, then design an algorithm or mechanism whose behavior can be analyzed.

Why advanced robotics needs rigorous mathematics

Robots move in space, and space has structure. A rigid body does not move like a point on a line; it translates and rotates. Rotations cannot be handled safely by treating roll, pitch, and yaw as ordinary independent coordinates in every situation, because orientation has a non-Euclidean geometry. Modern robot kinematics therefore uses matrices, homogeneous transformations, twists, exponentials, and Lie groups such as \(SO(3)\) for rotations and \(SE(3)\) for rigid-body motions (Murray, Li, and Sastry, 1994; Lynch and Park, 2017).

A configuration is a complete specification of the robot’s geometric position within its allowable motion variables. For a simple revolute-joint arm, the configuration may be the vector of joint angles

\[ q = \begin{bmatrix} q_1 & q_2 & \cdots & q_n \end{bmatrix}^\top . \]

The set of all possible configurations is called the configuration space. If a planar two-link arm has two revolute joints, its configuration space is often modeled as a torus-like space because each joint angle wraps around periodically. This geometric fact affects collision checking, interpolation, planning, and singularity analysis.

A model is a mathematical representation of a physical system. In robotics, models may describe geometry, velocity, force, probability, contact, computation time, communication delay, sensor noise, or learned behavior. A model is not valuable because it is perfectly complete; it is valuable when it captures the structure needed for prediction, design, or explanation.

For example, the dynamic model of a rigid-link manipulator is often written in the form

\[ M(q)\ddot q + C(q,\dot q)\dot q + g(q) = \tau , \]

where \(q\) is the vector of generalized coordinates, \(M(q)\) is the inertia matrix, \(C(q,\dot q)\dot q\) represents velocity-dependent terms such as Coriolis and centrifugal effects, \(g(q)\) represents gravity terms, and \(\tau\) is the vector of generalized actuator forces or torques. This equation is not just notation; it expresses how geometry, mass distribution, velocity, gravity, and actuation jointly determine motion. Standard robot dynamics texts derive this structure using Newton–Euler, Lagrangian, and related formulations (Spong, Hutchinson, and Vidyasagar, 2006; Siciliano et al., 2009).

The graduate-level challenge is not merely to memorize such equations. It is to understand what assumptions make them valid, how they change under coordinates, how numerical algorithms use them, and how real hardware violates ideal assumptions through friction, backlash, saturation, compliance, delay, sensor noise, and unmodeled contact.

Autonomy is not magic; it is structured decision-making under constraints

The word autonomy means the ability of a system to perform tasks with limited human intervention. It does not mean unlimited intelligence, perfect understanding, or freedom from engineering constraints. A robot may be autonomous for one task and not autonomous for another. A warehouse robot may navigate safely in a mapped building but fail outdoors. A surgical robot may have excellent precision but still require direct human supervision. A planetary rover may plan local motion autonomously while receiving strategic goals from Earth.

Autonomy requires decisions. A decision may be as small as choosing the next motor torque or as large as selecting which object to inspect first. These decisions are constrained by physical laws, sensor limitations, computation time, safety requirements, and task goals.

Three important terms appear throughout the book:

Control is the design of input commands that influence a system’s behavior. For example, a joint-space controller may compute motor torques so that a robot arm follows a desired joint trajectory. Classical robot control includes PID control, computed-torque control, inverse dynamics, adaptive control, robust control, and operational-space control (Craig, 2005; Spong, Hutchinson, and Vidyasagar, 2006).

Planning is the computation of a feasible or desirable path, trajectory, or sequence of actions before or during execution. For example, a mobile robot may plan a collision-free path through a building, or a manipulator may plan how to move an object without hitting obstacles. Configuration-space planning, graph search, sampling-based planning, and kinodynamic planning are central methods in the planning literature (LaValle, 2006).

Estimation is the inference of hidden or uncertain quantities from measurements and models. For example, a drone may estimate its pose from inertial measurements and camera images; a mobile robot may estimate both its location and a map of the environment. Bayesian filtering, Kalman filtering, particle filtering, factor graphs, and smoothing are standard estimation tools in robotics (Thrun, Burgard, and Fox, 2005).

These are different ideas, but they are rarely isolated in real robots. A robot that plans without estimation may plan from a false world model. A controller without planning may be locally stable but globally useless. A perception system without control relevance may recognize objects but fail to help the robot manipulate them.

An example: picking up a cup is a research-level problem

A human can pick up a cup so easily that the task seems simple. For a robot, the task crosses nearly the entire discipline.

First, the robot must perceive the cup. If it uses a camera, image pixels must be converted into geometric and semantic information: where the cup is, which pixels belong to it, what shape it has, whether it is transparent or reflective, and whether its pose estimate is uncertain.

Second, the robot must reason about kinematics. It must know whether its hand can reach the cup, which joint angles produce the desired end-effector pose, and whether the motion passes near a singularity. A singularity is a configuration where the robot loses the ability to produce certain instantaneous motions or where small task-space motions may require very large joint velocities.

Third, the robot must plan a collision-free motion. It must avoid the table, nearby objects, its own body, and possibly humans. If the environment changes, the plan may need to be updated.

Fourth, the robot must handle contact. Grasping is not just placing fingers near an object. The robot must create contact forces that resist gravity and disturbances while respecting friction limits. A friction cone describes the set of contact force directions that can be supported without slipping. If the commanded force lies outside this cone, the object may slide.

Fifth, the robot must execute the motion using control. Motors have torque limits, transmissions have compliance, sensors have noise, and delays exist. A controller must remain stable and accurate despite these effects.

Finally, the robot may use learning. It may learn object affordances, predict grasp success, adapt to surface friction, or improve visuomotor policies from data. But learning does not remove the need for mechanics, estimation, and control; it adds another layer of modeling and uncertainty.

This small example explains why advanced robotic engineering requires both theory and experiment. The theory tells us what structure to exploit. The experiment tells us which assumptions survived contact with hardware.

What “advanced” means in this book

In this book, “advanced” does not mean obscure. It means that ideas are treated with the level of precision needed for research and high-consequence engineering.

An introductory robotics course may teach that a robot arm has joints and that forward kinematics maps joint angles to end-effector pose. Here we go further: we examine coordinate frames, representation singularities, Lie group structure, spatial velocity, body velocity, adjoint transformations, manipulator Jacobians, manipulability, constraints, and numerical conditioning.

An introductory control course may teach that feedback reduces error. Here we go further: we analyze nonlinear stability, Lyapunov functions, passivity, operational-space dynamics, optimal control, model predictive control, control Lyapunov functions, and control barrier functions.

An introductory artificial intelligence course may teach search and machine learning algorithms. Here we ask which of those algorithms remain reliable when attached to moving hardware with limited sensing, limited computation, uncertain dynamics, and safety constraints.

The book’s goal is not to make every reader a specialist in every subfield. Instead, it builds a coherent map: enough depth to read research papers critically, design experiments carefully, understand assumptions, and connect theory to robotic systems.

How the chapters fit together

The book begins with mathematical foundations because robotics is not a collection of independent tricks. Linear algebra describes vectors, transformations, projections, and least-squares estimation. Differential equations describe motion. Optimization describes planning, estimation, calibration, and control. Probability describes uncertainty. Graph theory describes maps, factor graphs, task networks, and multi-robot communication. Numerical conditioning explains why a mathematically correct formula may still fail in computation.

Rigid-body motion comes next because every robot must move through space. Rotations, transformations, twists, screws, and Lie groups provide the language for modern kinematics, estimation, and control.

Kinematics then asks: given joint variables, where is the robot? Given a desired pose, what joint variables could produce it? Velocity kinematics and Jacobians extend this question to instantaneous motion, force relationships, singularities, and redundancy.

Dynamics asks why motion changes. Mass, inertia, gravity, actuator limits, constraints, and contact forces determine what motions are physically possible. Control then uses models and feedback to make desired behavior occur.

State estimation, perception, localization, and mapping address the fact that robots do not have direct access to the true world state. They must infer it from noisy, incomplete, and sometimes ambiguous measurements.

Motion planning and trajectory generation address the question of how to move through the world while respecting geometry, dynamics, and constraints.

Contact, grasping, manipulation, and specialized robot domains show how difficult the world becomes when robots leave clean laboratory assumptions. Legged robots, aerial robots, marine robots, field robots, and planetary rovers face different physics and deployment constraints.

Human–robot interaction, multi-robot systems, safety, verification, reliability, ethics, and research methods complete the engineering picture. Robots operate among people, institutions, regulations, and social consequences. A technically impressive robot can still be unsafe, irreproducible, ethically harmful, or scientifically unconvincing if it is not evaluated carefully.

The role of learning in modern robotics

Machine learning has become an important part of robotics, especially in perception, grasp prediction, policy learning, representation learning, and sim-to-real transfer. However, the central lesson of this book is that robot learning should be understood as part of robotic systems engineering, not as a replacement for it.

A learned perception module may estimate object pose, but the robot still needs calibration, uncertainty propagation, and control. A learned policy may output actions, but the actions still pass through actuators, contact dynamics, and safety constraints. A learned dynamics model may improve planning, but it must be evaluated against physical data and failure cases.

This distinction is important: learning a component is different from learning complete robotic behavior. The first may improve a perception pipeline, controller, planner, or model. The second attempts to produce integrated behavior directly from data. Both approaches can be useful, but they have different data requirements, safety risks, interpretability properties, and validation burdens.

Throughout the later chapters, learning is treated with the same engineering discipline as any other modeling tool: What data were used? What distribution shift is expected? What uncertainty is represented? What constraints are guaranteed? What happens under sensor failure, actuator saturation, unexpected contact, or human intervention?

A disciplined way to think like a roboticist

When approaching any robotic system, ask six questions.

First, what is the task? A vague goal such as “make the robot intelligent” is not yet an engineering problem. A better task description is: “Move a 2 kg object from a table to a shelf within 10 seconds without exceeding contact force limits or colliding with nearby objects.”

Second, what is the system? Specify the robot morphology, actuators, sensors, computational hardware, environment, and interaction assumptions.

Third, what is the state and what is uncertain? Decide which variables matter and which are unknown. For example, the robot may know its joint angles accurately but only estimate the object’s pose approximately.

Fourth, what models are being used? These may include rigid-body dynamics, camera projection, wheel-ground contact, friction, communication delay, or learned prediction models.

Fifth, what constraints must never be violated? Safety, torque limits, collision avoidance, thermal limits, legal restrictions, and human comfort may be more important than nominal performance.

Sixth, how will success be measured? A research claim requires evidence. That evidence may include mathematical proof, simulation, hardware experiments, benchmark comparison, statistical analysis, ablation studies, and reproducible implementation details.

These questions prevent a common failure mode in robotics: solving a beautiful subproblem while ignoring the system that must actually work.

What you should expect from this book

This book assumes that you are willing to engage with mathematics, algorithms, and physical interpretation together. It will not hide equations, but it will explain why the equations exist. It will not treat hardware as an afterthought, but it will also not reduce robotics to mechanical construction. It will introduce modern research tools, but it will continually ask what assumptions they require and how they behave on real systems.

When you see a formula, ask what physical or computational relationship it expresses. When you see an algorithm, ask what information it needs and what can go wrong. When you see an experimental result, ask what was measured, what was controlled, what was uncertain, and whether the conclusion follows from the evidence.

Robotics is difficult because the world is difficult. Bodies have inertia. Surfaces have friction. Sensors are noisy. Computation takes time. Humans are variable. Environments change. Yet these difficulties are also what make robotics intellectually rich. A successful robot is a negotiated agreement among geometry, physics, information, computation, design, and purpose.

That agreement is the subject of this book.

References

Craig, J. J. (2005). Introduction to Robotics: Mechanics and Control (3rd ed.). Pearson Prentice Hall.

ISO 8373:2021. (2021). Robotics — Vocabulary. International Organization for Standardization.

LaValle, S. M. (2006). Planning Algorithms. Cambridge University Press.

Lynch, K. M., & Park, F. C. (2017). Modern Robotics: Mechanics, Planning, and Control. Cambridge University Press.

Murray, R. M., Li, Z., & Sastry, S. S. (1994). A Mathematical Introduction to Robotic Manipulation. CRC Press.

Siciliano, B., & Khatib, O. (Eds.). (2016). Springer Handbook of Robotics (2nd ed.). Springer.

Siciliano, B., Sciavicco, L., Villani, L., & Oriolo, G. (2009). Robotics: Modelling, Planning and Control. Springer.

Spong, M. W., Hutchinson, S., & Vidyasagar, M. (2006). Robot Modeling and Control. Wiley.

Thrun, S., Burgard, W., & Fox, D. (2005). Probabilistic Robotics. MIT Press.