Loading field map
Robot kinematics and dynamics is the branch of robotics concerned with the mathematical description of a robot's motion and the forces that cause it. It provides the formal language used to answer two fundamental questions: Where can the robot go, and how does it get there? and What forces and torques must its actuators produce to move it in a desired way? The field is not a single unified theory but a structured collection of mathematical frameworks, each developed to address a different aspect of these questions. Understanding the field requires seeing how these frameworks fit together, what each one can and cannot do, and how they jointly form the foundation for robot control, simulation, and design.
The subject splits naturally along the line separating geometry from physics. Kinematics is the study of motion without regard to its causes. It describes positions, velocities, and accelerations of a robot's parts as pure geometry in time. Dynamics adds forces, torques, masses, and inertia, explaining why the motion occurs. Both are further divided into forward and inverse problems.
Forward kinematics asks: given the angles or positions of the robot's joints, where is the end effector (the hand, tool, or sensor at the end of the kinematic chain)? This is a direct geometric calculation. For a serial robot—one built as a chain of links connected by joints, like a human arm—the answer is obtained by multiplying together a sequence of transformation matrices, each describing the position and orientation of one link relative to the previous one. The standard formalism for this is the Denavit–Hartenberg convention, which assigns coordinate frames to each link in a systematic way so that the entire chain can be described by a small set of parameters.
Inverse kinematics is the reverse problem: given a desired position and orientation of the end effector, what joint angles produce it? This is fundamentally harder. The forward map from joint space to end-effector space is nonlinear, and its inverse may have multiple solutions, a unique solution, or no solution at all, depending on the robot's geometry and the target point. A robot with six revolute joints, for example, can generally reach a given pose in a finite number of distinct configurations, but a robot with five joints cannot reach arbitrary poses in space. Inverse kinematics is typically solved either by closed-form analytic expressions for special geometries (such as robots with intersecting joint axes) or by iterative numerical methods that converge to a solution from an initial guess. The existence of multiple solutions is not a nuisance but a design feature: it allows the robot to avoid obstacles or keep its joints within limits by choosing among alternative configurations.
Differential kinematics bridges the gap between kinematics and dynamics. It relates joint velocities to end-effector velocities through the Jacobian matrix, a linear map that depends on the current configuration. The Jacobian is central to the field because it also encodes kinematic singularities—configurations where the robot loses the ability to move in certain directions, no matter how fast its joints turn. At a singularity, the Jacobian loses rank, and the robot's instantaneous mobility is reduced. Singularities are not merely mathematical curiosities; they cause practical problems in control, as they can lead to very large joint velocities for modest end-effector motions and can make the robot's behavior unpredictable.
Forward dynamics takes the forces and torques applied at the joints, along with the current state (positions and velocities), and computes the resulting accelerations. This is the problem used in simulation: given what the actuators are doing, where will the robot be a moment from now? Inverse dynamics is the opposite: given a desired trajectory of positions, velocities, and accelerations, what joint torques are required to produce it? Inverse dynamics is the basis of computed-torque control, a technique in which the controller cancels the robot's nonlinear dynamics and substitutes a desired linear behavior.
Dynamics can be derived in two mathematically equivalent but practically distinct ways. The Newton–Euler formulation applies Newton's second law and its rotational analogue to each link in turn, propagating forces and torques from one link to the next. It yields a set of recursive equations that are computationally efficient—the standard algorithm runs in time linear in the number of joints—and is the method of choice for real-time control and simulation. Its conceptual cost is that it produces a large number of intermediate variables (constraint forces at each joint) that must be eliminated to obtain the compact equations of motion.
The Lagrangian formulation takes a global view. It defines the robot's kinetic and potential energy as functions of the joint coordinates, then applies the Euler–Lagrange equation to derive the equations of motion directly. The result is a single matrix equation of the form
\[ M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau \]
where \(M\) is the mass matrix, \(C\) captures Coriolis and centrifugal effects, \(G\) is the gravity term, and \(\tau\) is the vector of joint torques. This form is highly revealing: it shows that the dynamics are nonlinear, that the effective inertia seen at each joint depends on the configuration, and that motions of one joint induce forces on others through the coupling terms. The Lagrangian approach is preferred for analysis, controller design, and deriving properties such as passivity and energy conservation, even though its direct computational cost is higher than the recursive Newton–Euler method.
The two formulations are not rival theories but complementary tools. Modern practice often uses the recursive Newton–Euler algorithm for computation and the Lagrangian form for insight. The relationship between them is well understood: the Lagrangian equations can be derived from the Newton–Euler recursion by eliminating the internal constraint forces, and the recursive algorithm can be seen as an efficient way to evaluate the Lagrangian equations without ever forming the full matrices.
The discussion so far assumes a serial chain, where each joint connects exactly two links and the links form a single open path. Many robots, however, contain closed loops. A parallel robot—such as the Stewart platform used in flight simulators, or a delta robot used in high-speed pick-and-place—consists of multiple kinematic chains connecting a fixed base to a moving platform. The platform's motion is constrained by all chains simultaneously, and the kinematics and dynamics become substantially more complex.
For parallel robots, the forward kinematics problem (given actuator lengths, find the platform pose) is generally much harder than the inverse problem, which is the reverse of the serial case. The dynamics must account for constraint forces that maintain the closed loops, and the equations of motion become a set of differential-algebraic equations rather than ordinary differential equations. The standard approach is to introduce Lagrange multipliers for the constraints, or to use the Newton–Euler method with explicit constraint forces. Parallel robots also exhibit a richer singularity structure, including configurations where the platform gains or loses degrees of freedom in ways that have no analogue in serial chains.
The study of constrained systems connects robot dynamics to the broader field of multibody dynamics, where the same mathematical machinery is used for mechanisms, vehicles, and biomechanical systems. The key conceptual shift is that the configuration space is no longer a simple product of joint variables but a manifold defined by the constraint equations.
A distinct and historically important approach to kinematics uses screw theory, which represents a rigid-body motion as a rotation about an axis combined with a translation along that same axis. A screw is a six-dimensional vector that encodes both the axis and the pitch (the ratio of translation to rotation). The set of all possible instantaneous motions of a robot is a subspace of the six-dimensional space of screws, and the set of all possible forces and torques is its dual.
This geometric viewpoint, developed in the nineteenth century and revived in robotics in the late twentieth century, offers several advantages. It treats rotations and translations in a unified way, avoiding the singularities that plague Euler-angle representations. It provides elegant characterizations of singularities and of the conditions under which a robot can or cannot exert forces in given directions. It also leads to the product of exponentials formula for forward kinematics, which expresses the end-effector pose as a product of matrix exponentials, one per joint, in a way that is coordinate-free and often more transparent than the Denavit–Hartenberg approach.
Screw theory is not a separate subfield so much as a different mathematical language for the same content. Its practitioners argue that it reveals structure that the coordinate-based approach obscures, particularly for analysis of singularities, mobility, and force transmission. The two languages coexist in the literature, and many modern textbooks present both, noting that the choice is largely one of taste and problem type.
Contemporary robot kinematics and dynamics is a mature field with a stable core of results, but it continues to evolve in response to new hardware and new applications. Several trends define the current landscape.
Computational efficiency remains a live concern. The recursive Newton–Euler algorithm, developed in the 1980s, is near-optimal for serial chains, but modern robots with many degrees of freedom—humanoids, quadrupedal robots, multi-fingered hands—push the limits of real-time computation. Research continues on algorithms that exploit sparsity, parallel computing, and the special structure of branched or closed-loop mechanisms.
Contact and impact have moved from the periphery to the center of the field. Traditional robot dynamics assumed free motion in space, but legged robots, manipulation, and physical human–robot interaction all involve sustained or intermittent contact with the environment. Modeling contact requires combining rigid-body dynamics with friction laws, impact models, and constraint stabilization, and the resulting hybrid systems (switching between free motion and constrained motion) pose challenges for both simulation and control. This area draws on classical mechanics but extends it in ways that are still not fully settled, particularly for impacts and for systems with many simultaneous contacts.
Uncertainty and identification are increasingly important. The equations of motion contain parameters—masses, inertias, centers of mass, friction coefficients—that are never known exactly. The field of dynamic parameter identification addresses how to estimate these parameters from measured motion and torque data, and how to design excitation trajectories that make the estimates well-conditioned. This is not a peripheral concern but a practical necessity: model-based controllers are only as good as their parameter estimates, and the gap between nominal and actual parameters is a major source of performance degradation.
The relationship to control has shifted. In the early decades, kinematics and dynamics were seen as prerequisites for control: first model the robot, then design a controller for the model. Modern practice increasingly treats modeling and control as joint problems. Techniques such as operational-space control formulate the dynamics directly in the space of the task (the end-effector coordinates) rather than in joint space, and impedance control deliberately shapes the apparent dynamics of the robot to achieve compliant behavior. These approaches do not replace the classical formulations but build on them, using the same equations in different coordinate systems.
Soft and continuum robots challenge the assumptions of the classical theory. A soft robot has no discrete joints and links; it deforms continuously, and its configuration space is infinite-dimensional. The field has responded by developing reduced-order models—often based on the theory of Cosserat rods, which describe a slender body by its centerline and cross-section orientation—that approximate the continuum with a finite set of modes. These models are far less mature than the rigid-body theory, and many fundamental questions about their accuracy, computational cost, and suitability for control remain open.
Throughout these developments, the core intellectual content of the field has remained stable. The kinematics of rigid bodies, the Jacobian, the equations of motion, and the distinction between forward and inverse problems are as central today as they were when the field was consolidated in the 1970s and 1980s. What has changed is the range of systems to which these tools are applied, the computational scale at which they are used, and the depth of integration with sensing, estimation, and control. The field is best understood not as a sequence of competing paradigms but as a coherent mathematical toolkit whose components—geometric, analytic, recursive, and computational—are selected and combined according to the problem at hand.