Product of exponentials and screw theory

The Denavit–Hartenberg method needs a frame on every link, placed by careful rules. In 1984 Roger Brockett showed a different route to forward kinematics that needs only two things: the robot's tool pose at its zero configuration, and a description of each joint axis in a single fixed frame. The resulting product of exponentials (PoE) formula is built on screw theory, a nineteenth-century description of rigid motion that has become central to modern robotics. PoE is easy to set up from a CAD model, handles revolute and prismatic joints uniformly, behaves well when joint axes are nearly parallel, and leads directly to the Jacobian of lesson 10. This lesson develops the ideas from the ground up.

Rotations as exponentials

Lesson 3 introduced exponential coordinates: a rotation of angle \( \theta \) about unit axis \( \hat{\boldsymbol{\omega}} \) is captured by the vector \( \hat{\boldsymbol{\omega}}\theta \). The connection to the exponential function comes from angular velocity. A body spinning at constant angular velocity \( \hat{\boldsymbol{\omega}} \) has orientation obeying \( \dot{R} = [\hat{\boldsymbol{\omega}}]R \), where \( [\hat{\boldsymbol{\omega}}] \) is the skew-symmetric matrix of lesson 3. Like the scalar equation \( \dot{x} = ax \), this differential equation is solved by an exponential, here a matrix exponential:

The right-hand side is Rodrigues' formula again. The infinite series \( e^{A} = I + A + A^2/2! + \cdots \) collapses to three terms because the skew matrix of a unit vector satisfies \( [\hat{\boldsymbol{\omega}}]^3 = -[\hat{\boldsymbol{\omega}}] \), so the powers cycle and the sine and cosine series appear. The matrix logarithm goes the other way, from a rotation matrix back to \( \hat{\boldsymbol{\omega}}\theta \). Exponentials and logarithms are how modern software integrates gyroscope readings, interpolates orientations and measures orientation errors.

Twists and screw motions

Rigid bodies do more than rotate: they translate too. Chasles' theorem, proved in 1830, states that every rigid-body displacement can be achieved by a screw motion: a rotation about some line in space combined with a translation along the same line, like a nut turning on a bolt. The ratio of translation to rotation is the pitch of the screw. A pure rotation has zero pitch; a pure translation has infinite pitch.

A screw axis is described by a six-vector \( \mathcal{S} = (\boldsymbol{\omega}, \mathbf{v}) \). For a revolute joint, \( \boldsymbol{\omega} \) is the unit vector along the joint axis and \( \mathbf{v} = -\boldsymbol{\omega} \times \mathbf{q} \), where \( \mathbf{q} \) is any point on the axis. Physically, \( \mathbf{v} \) is the velocity that the point at the origin would have if the body rotated about the axis at unit speed. For a prismatic joint, \( \boldsymbol{\omega} = \mathbf{0} \) and \( \mathbf{v} \) is the unit vector along the direction of sliding. A screw axis scaled by a rate is a twist, the six-dimensional velocity of a rigid body, combining angular and linear velocity in one object.

The exponential of a twist

Just as a rotation is the exponential of a skew matrix, a rigid motion is the exponential of a twist. Write the twist as a 4 × 4 matrix:

Then moving along the screw axis by \( \theta \), an angle for revolute joints or a distance for prismatic ones, produces the homogeneous transform:

For a prismatic joint, with \( \boldsymbol{\omega} = \mathbf{0} \), this reduces to a pure translation by \( \mathbf{v}\theta \). The formula is exact and cheap: one Rodrigues rotation plus a 3 × 3 matrix times a vector. It is the motion a body would undergo if it moved at the constant twist \( \mathcal{S} \) for time \( \theta \).

The product of exponentials formula

Now consider a serial arm. Place the arm in its zero configuration, all joint variables zero, and let M be the tool pose in the fixed base frame in that configuration. Express each joint's screw axis \( \mathcal{S}_i \) in the base frame, also at the zero configuration. Then the forward kinematics for any joint values is:

The reasoning is elegant. Moving only the last joint by \( \theta_n \) moves the tool by \( e^{[\mathcal{S}_n]\theta_n} \) about joint n's axis, which is still where it was at zero. Then moving joint n − 1 carries joint n and the tool together about axis n − 1, which also has not moved, and so on back to the base. Because each exponential is applied about an axis that has not yet been displaced, every screw axis can be written once, at the zero configuration, in the base frame. No link frames are needed at all.

There is also a body form, in which the screw axes \( \mathcal{B}_i \) are expressed in the tool frame at zero and M appears on the left: \( T(\boldsymbol{\theta}) = M e^{[\mathcal{B}_1]\theta_1} \cdots e^{[\mathcal{B}_n]\theta_n} \). The two forms are related by the adjoint map, \( \mathcal{B}_i = [\mathrm{Ad}_{M^{-1}}]\mathcal{S}_i \), which transforms twists between frames. The space form is convenient for analysis in the base frame; the body form suits tool-centered tasks and leads to the body Jacobian.

Worked example: the planar 3R arm again

Take the planar 3R arm of lesson 5 with link lengths \( L_1 \), \( L_2 \) and \( L_3 \). In the zero configuration the arm lies along the x-axis, so the tool frame has identity orientation and sits at \( (L_1 + L_2 + L_3, 0, 0) \). All joint axes point out of the plane, along z, and pass through \( (0, 0, 0) \), \( (L_1, 0, 0) \) and \( (L_1 + L_2, 0, 0) \). Using \( \mathbf{v} = -\boldsymbol{\omega} \times \mathbf{q} \):

That is the whole model: one matrix M and three six-vectors, read straight from a drawing of the arm at zero. Evaluating the product of exponentials reproduces the DH result exactly, as the code below confirms for the same joint angles as lesson 5.

import numpy as np
np.set_printoptions(precision=4, suppress=True)
def skew(w): return np.array([[0,-w[2],w[1]],[w[2],0,-w[0]],[-w[1],w[0],0]])
def exp6(S, th):
    w, v = np.array(S[:3], float), np.array(S[3:], float)
    T = np.eye(4)
    if np.allclose(w, 0):                      # prismatic
        T[:3,3] = v*th; return T
    W = skew(w)
    T[:3,:3] = np.eye(3) + np.sin(th)*W + (1-np.cos(th))*W@W
    T[:3,3] = (np.eye(3)*th + (1-np.cos(th))*W + (th-np.sin(th))*W@W) @ v
    return T
def poe(Slist, M, thetas):
    T = np.eye(4)
    for S, th in zip(Slist, thetas): T = T @ exp6(S, th)
    return T @ M
L1, L2, L3 = 0.4, 0.3, 0.1
M = np.eye(4); M[0,3] = L1 + L2 + L3
S = [(0,0,1, 0,0,0), (0,0,1, 0,-L1,0), (0,0,1, 0,-(L1+L2),0)]
T = poe(S, M, np.radians([30, 45, -60]))
print("tool position:", T[:3,3], " heading:", round(np.degrees(np.arctan2(T[1,0], T[0,0])), 1), "deg")

tool position: [0.5206 0.5157 0.    ]  heading: 15.0 deg

The result, about (0.521, 0.516) m with a 15° heading, matches the DH calculation of lesson 5, as it must: the two methods are different parameterizations of the same geometry.

A spatial example with a prismatic joint

Consider an RRP arm: a vertical waist joint at the origin, a horizontal shoulder joint 0.5 m above the base, and a prismatic joint that extends a boom horizontally. In the zero configuration the boom points along x, and the tool sits at the end of a 0.6 m boom at \( (0.6, 0, 0.5) \).

Why use PoE?

PoE has practical advantages over DH. It needs no link frames, so it can be set up directly from a CAD model or a URDF by reading joint axes and positions. Revolute and prismatic joints are treated identically. Small changes in the geometry produce small changes in the screw axes, even when axes are nearly parallel, which makes PoE well suited to kinematic calibration. And the screw axes are themselves the columns of the space Jacobian, as lesson 10 shows, so velocity kinematics comes almost for free. The textbook Modern Robotics by Lynch and Park is built entirely on this approach, and its free software library implements the functions used here. DH remains common in data sheets and older controllers, so practicing engineers should be comfortable with both and able to convert between them.