Loading field map
A robot arm must know where its hand is, how fast it is moving, and what forces it exerts on the world. These three questions—geometry, velocity, and dynamics—define the subfield of robot kinematics and dynamics. Since the 1950s, researchers have developed nine major methodological frameworks to answer them. The history is not a smooth march toward a single best method; it is a series of trade-offs between algebraic convenience, geometric insight, computational speed, and control relevance. Each framework made a distinctive commitment about what mattered most, and those commitments still divide the field today.
Before 1955, engineers described robot links with ad-hoc geometry. The Denavit-Hartenberg (DH) Kinematics framework changed that by introducing a systematic parameterization: four numbers per joint (link length, twist, offset, and joint angle) that completely describe the relative position and orientation of consecutive links. DH turned the forward kinematics problem—given joint angles, find the end-effector pose—into a routine matrix multiplication. This standardization was the infrastructure on which nearly everything else was built.
Analytic Inverse Kinematics (1968) addressed the harder reverse problem: given a desired end-effector pose, find the joint angles that achieve it. For many common arm geometries, closed-form solutions exist, and these rely on the DH parameterization to set up the trigonometric equations. Where DH provided a universal forward language, analytic inverse kinematics exploited the specific structure of that language to solve the inverse problem symbolically. The two frameworks remain tightly coupled in practice: DH is the standard input format, and analytic solutions are still preferred when available because they are fast and predictable.
By the late 1960s, controlling a manipulator in real time required more than static poses. Jacobian-Based Differential Kinematics (1969) introduced the Jacobian matrix, which maps joint velocities to end-effector velocities. This made resolved-rate control possible: instead of computing inverse kinematics at every timestep, a controller could use the Jacobian inverse to convert a desired Cartesian velocity into joint velocity commands. The Jacobian also gave access to the manipulator's singularities—configurations where the arm loses a degree of freedom—and became the standard tool for velocity-level analysis.
Screw-Theoretic Kinematics (1978) offered a fundamentally different worldview. Where the Jacobian works in coordinates, screw theory treats motion as instantaneous twists about axes in space, using a geometric language that does not depend on a particular reference frame. A screw describes both rotation and translation along a single axis, and the set of all possible twists forms a Lie algebra. This framework absorbed the Jacobian's velocity mapping as a special case but reframed it in coordinate-free terms. The tension between the algebraic Jacobian and the geometric screw view persists: the Jacobian is easier to compute and implement, while screw theory reveals structural properties—such as reciprocity and instantaneous mobility—that are invisible in coordinate form.
Product-of-Exponentials (POE) Kinematics (1983) took the screw-theoretic insight and formalized it using Lie groups. Instead of DH's four parameters per joint, POE represents each joint as an exponential of a twist: a single matrix that encodes both the joint axis and the motion it produces. The forward kinematics becomes a product of exponentials, each corresponding to a joint's screw motion. POE has several technical advantages over DH: it requires fewer parameters for many common joints, it handles prismatic and revolute joints uniformly, and it avoids the singularities that DH encounters when consecutive joint axes are parallel. More importantly, POE is a global representation—it describes the entire motion of the arm in a single coordinate frame, whereas DH builds a chain of local frames. This geometric unification made POE the dominant framework in modern robotics research and textbooks, especially for modeling, calibration, and control.
Kinematics describes motion without forces; dynamics adds the torques and forces that cause that motion. The first systematic dynamics framework was Lagrangian Manipulator Dynamics (1980), which derives the equations of motion from the kinetic and potential energy of the arm. The Lagrangian approach is theoretically elegant: it produces closed-form symbolic equations that reveal the structure of inertia, Coriolis, centrifugal, and gravity terms. Its cost, however, grows as O(n⁴) with the number of joints, making it impractical for real-time control of arms with more than a few degrees of freedom.
Recursive Newton-Euler Dynamics (1980) took the opposite philosophical stance. Instead of starting from energy, it propagates forces and accelerations link by link using Newton's second law and Euler's equation for rotation. The algorithm runs in O(n) time—linear in the number of joints—and is the standard workhorse for inverse dynamics in real-time control. Where the Lagrangian framework commits to symbolic clarity, the Recursive Newton-Euler framework commits to computational efficiency. The two are not in direct competition: Lagrangian dynamics is used for analysis, simulation, and model-based control design, while Recursive Newton-Euler is used when the dynamics must be computed at kilohertz rates on embedded hardware.
Both the Lagrangian and Recursive Newton-Euler frameworks assume a fixed-base serial chain—a single arm rooted in the world. Articulated-Body Dynamics Algorithms (1983) extended dynamics to tree-structured robots (multiple arms, legs, or fingers) and floating-base systems (humanoids, space robots). The key idea is the articulated-body inertia: instead of computing forces recursively from the base outward, the algorithm computes the effective inertia of each subtree as seen from its parent joint. This allows forward dynamics—given torques, find accelerations—to be computed in O(n) time for tree structures, and it handles closed kinematic chains through constraint projection. Articulated-body algorithms absorbed the recursive efficiency of Newton-Euler while generalizing to the topologies that real robots actually have.
By the mid-1980s, kinematics and dynamics had been thoroughly developed in joint space. But a robot's task—moving a tool, applying a force—is defined in the end-effector's Cartesian space. Operational-Space Dynamics (1987) reformulated the dynamics problem in task coordinates using the Jacobian. The framework produces an operational-space inertia matrix that decouples the dynamics of the end-effector from the rest of the arm, enabling simultaneous control of motion and force in the task frame. This was a synthesis: it used the Jacobian (from differential kinematics) to project joint-space dynamics into task space, and it relied on the Recursive Newton-Euler algorithm for efficient computation of the required torques. Operational-space dynamics became the foundation for impedance control, force control, and whole-body control in humanoids.
Today, two frameworks dominate the field, each with a different philosophical commitment. Product-of-Exponentials Kinematics is the leading framework for research and teaching because of its geometric elegance, global representation, and seamless integration with modern control theory. Recursive Newton-Euler Dynamics is the leading framework for real-time control because of its O(n) efficiency and straightforward implementation on embedded hardware. They coexist in a productive division of labor: POE models the arm, and RNE computes the torques.
What the leading frameworks agree on is that a unified geometric representation (screw theory, Lie groups) is the best way to think about kinematics, and that recursive algorithms are the only practical way to compute dynamics in real time. What they disagree on is whether the geometric abstraction is worth the computational overhead. POE requires exponentiating twists, which is more expensive than DH's simple matrix multiplications. For high-speed control loops, DH and the Jacobian remain faster, and many industrial robot controllers still use them.
A deeper debate concerns the role of learning. Classical frameworks like DH, POE, and Recursive Newton-Euler assume a known kinematic and dynamic model. Learning-based approaches can identify model parameters from data or bypass the model entirely. The tension is not a rejection of the classical frameworks—learning methods often use DH or POE as the model structure and only learn the parameters—but it raises the question of whether the geometric and analytic commitments of the classical frameworks will remain central as data-driven methods improve. For now, the classical frameworks provide the structure that makes learning efficient, and the two approaches are increasingly integrated rather than opposed.