[Intermediate] — Given the joint angles, where is the hand? The easy direction, and the notation the field standardised on.
The problem
This direction is always solvable, always unique, and has a closed form. Multiply the link transforms together and you’re done:
That’s the whole idea. Everything else in this note is bookkeeping to make the individual link transforms systematic.
Contrast with inverse kinematics, which may have no solution, one, several, or infinitely many. Forward is easy; inverse is the subject.
A worked two-link arm
Before the general machinery, do it by hand — planar, two revolute joints, link lengths :
╱● (x, y)
╱ l₂
●───╱ θ₂
╱ l₁
╱ θ₁
●───────────
Note the . Joint angles are measured relative to the previous link, so they accumulate down the chain. That’s the essential structure of every serial manipulator, and it’s why the transforms multiply.
Do this one on paper. The general method is this, formalised — and the formalism is much easier to trust once you’ve seen the concrete version.
Denavit–Hartenberg parameters
The standard way to describe a serial chain. Four numbers per joint instead of a general six-DOF transform, achieved by constraining how you’re allowed to place the frames.
| Parameter | Meaning | Varies for |
|---|---|---|
| joint angle about | revolute joints | |
| link offset along | prismatic joints | |
| link length along | fixed | |
| link twist about | fixed |
The trick that makes four suffice: the frames are placed so that is the common normal between and . That constraint removes two of the six degrees of freedom, and what’s left is exactly four.
One parameter per joint varies — for a revolute joint, for a prismatic one. The other three are fixed by the robot’s geometry. So an -joint arm is described by a table, and that table is the robot as far as kinematics is concerned.
The link transform:
\cos\theta_i & -\sin\theta_i\cos\alpha_i & \sin\theta_i\sin\alpha_i & a_i\cos\theta_i \\ \sin\theta_i & \cos\theta_i\cos\alpha_i & -\cos\theta_i\sin\alpha_i & a_i\sin\theta_i \\ 0 & \sin\alpha_i & \cos\alpha_i & d_i \\ 0 & 0 & 0 & 1 \end{bmatrix}$$ **Don't memorise it.** It's $\text{Rot}_z(\theta)\,\text{Trans}_z(d)\,\text{Trans}_x(a)\,\text{Rot}_x(\alpha)$ multiplied out — recognising that decomposition is more useful than recalling the entries. ### The honest assessment **DH is the standard, it's what textbooks and papers use, and you need to be able to read it.** But: - **Two incompatible conventions exist** — "standard" (Denavit–Hartenberg 1955) and "modified" (Craig). They differ in whether the frame sits at the near or far end of the link, and **they produce different tables for the same robot.** Mixing them silently gives a wrong model. Always state which you're using - **Frame assignment is fiddly and error-prone**, especially with parallel or intersecting joint axes - **It cannot represent branching** — a humanoid or a hand isn't a serial chain - **It's degenerate when consecutive axes are parallel**, where the common normal isn't unique **What modern practice actually uses:** URDF (in ROS) and similar formats specify each joint with an explicit parent frame, child frame, origin transform and axis. **More verbose, no conventions to get wrong, handles trees, and generated directly from CAD.** → [[robotics/13-ros-and-robot-software|ROS and Robot Software]] > **Learn DH to read the literature. Use URDF to build the robot.** ## Common arm geometries Recognising these tells you a lot before you compute anything: **Articulated / anthropomorphic (6R)** — the classic industrial arm. Three joints position the wrist, three orient it. Large dexterous workspace. UR, KUKA, ABB, Fanuc. **SCARA** — two parallel revolute joints, one prismatic, one rotation. **4 DOF, planar motion plus vertical.** Fast, stiff vertically, cheap. Dominant in electronics assembly, and the reason is that pick-and-place doesn't need 6 DOF. **Cartesian / gantry** — three prismatic axes. **Kinematics is trivial** ($x, y, z$ *are* the joint values). 3D printers, CNC, large gantries. Stiff, accurate, big footprint for the workspace you get. **Delta** — a parallel robot, three arms to a common platform. **Extremely fast** because the motors are all on the fixed base and the moving mass is tiny. Small workspace. Used for high-speed picking. **Parallel arms invert the difficulty.** For a delta or a Stewart platform, **inverse kinematics is easy and forward kinematics is hard** — the opposite of a serial arm, and worth knowing so it doesn't surprise you. ## The spherical wrist The design decision that makes 6-DOF arms tractable. **If the last three joint axes intersect at a point**, then position and orientation **decouple**: - The first three joints determine the **wrist centre position** - The last three determine the **orientation** about it $$\mathbf{p}_{wrist} = \mathbf{p}_{tool} - d_6\,R\,\hat{z}$$ **That's why 6-DOF industrial arms have closed-form inverse kinematics** and arbitrary 6-joint arms generally don't. You solve a 3-DOF position problem, then a 3-DOF orientation problem, instead of one coupled 6-DOF problem. → [[robotics/06-inverse-kinematics|Inverse Kinematics]] **It's a mechanical design choice made specifically to make the maths solvable**, which is a nice example of hardware and software being co-designed. Some modern arms (notably some collaborative designs) give it up for other benefits, and pay in IK complexity. ## The tool frame Forward kinematics gives you the **flange** — the mounting face at the end of the last link. **That's not what you care about.** You care about the gripper tip, the welding point, the camera. $$^{0}_{tool}T = {^{0}_{flange}T}\ {^{flange}_{tool}T}$$ The **tool centre point (TCP)** offset is a fixed transform you supply. > **Getting the TCP wrong is one of the commonest practical errors**, and it's insidious: the robot is perfectly accurate and consistently wrong by a fixed offset. Worse, a TCP error in *orientation* produces a position error that changes as the wrist rotates, which looks like a much more complicated fault than it is. **Calibrate the TCP** rather than measuring it from CAD — the standard method is touching a single fixed point from four different orientations and solving for the offset that makes them agree. ## Where forward kinematics is used More places than "where's the hand": **Visualisation** — drawing the robot in RViz is forward kinematics at every joint. **Collision checking** — a planner needs every link's pose, for every candidate configuration, thousands of times per second. **This is why FK performance matters** even though a single evaluation is cheap. → [[robotics/10-motion-planning|Motion Planning]] **Inside inverse kinematics** — numerical IK solvers evaluate FK repeatedly to measure error and iterate. **Sensor placement** — a lidar on link 3 is somewhere in the world determined by joints 1–3. **Calibration** — comparing measured tool poses against FK-predicted ones is how you identify the true link lengths and joint offsets. ## Practical notes **Get link lengths from CAD, then verify physically.** Manufacturing tolerance and assembly are real. **A 1 mm link error and a 0.1° joint offset both matter** at the end of a metre-long arm. **Joint zero is a convention you must pin down.** Where is "zero" and which direction is positive? Nothing about the hardware tells you, and getting it wrong gives a mirrored robot. **Watch joint limits.** FK will happily compute a pose that requires a joint at 400°. The maths doesn't know about hard stops. **Cache along the chain.** Computing frames 1..n incrementally is much cheaper than recomputing each from the base — worth it inside a planner's inner loop. **Test against the physical robot.** Command a known configuration, measure where the tool actually is. **A tape measure has settled more kinematics arguments than any simulation.** --- ## Related - [[robotics/04-rigid-body-transforms|Rigid Body Transforms]] — the transforms being chained - [[robotics/06-inverse-kinematics|Inverse Kinematics]] — the hard direction - [[robotics/07-jacobians-and-singularities|Jacobians]] — differentiating this - [[robotics/index|Robotics map]]Software engineering
Infrastructure & security
Computer science
Hardware & engineering