Arm ⟷ IK

Why almost every industrial robot has a spherical wrist, and what changes when you add a seventh joint.

Line up six 6-DOF industrial arms from different decades and companies, a PUMA 560, an ABB IRB, a KUKA KR, a Stäubli, a Fanuc, a Yaskawa, and they share one feature: the last three rotation axes meet at a single point, exactly rather than approximately. This was never written into a standard. A 1968 PhD thesis proved that this is essentially the only 6-DOF geometry whose inverse kinematics a computer could solve in closed form, and designs that violated the criterion were dropped during productization because their IK could not run in real time on period hardware. The shape of the canonical robot arm is a computational artifact, and three results tell the story: Pieper (1968), Shimizu (2008), and Elias & Wen (2024).

1. What makes inverse kinematics hard

Forward kinematics asks: given the joint angles $q_i$, where is the end-effector? It is a straightforward product of transform matrices. Inverse kinematics (IK) asks the reverse: given a desired end-effector pose, which joint angles reach it? Even a planar 2-link arm shows the trouble. Many targets are unreachable, most reachable ones have two solutions (elbow up and elbow down), and the algebra involves arccos and atan2, so it is not polynomial.

At 6-DOF this becomes qualitatively harder. A pose in $\mathrm{SE}(3)$ has six numbers, three for position and three for orientation, so reaching an arbitrary pose needs at least six joints, and with exactly six the IK is six nonlinear equations in six unknowns. For a generic 6R manipulator with no special structure, this is equivalent to solving a degree-16 polynomial (Raghavan & Roth, 1990), which has no closed-form radical solution and must be solved numerically. Today that takes microseconds. In 1968 a single Newton iteration on a PDP-10 took tens to hundreds of milliseconds, while an industrial control loop needs at least 1 kHz, so closed-form IK was not merely preferable but necessary. The practical question of the era was which arm geometry would let the computer keep up.

2. Pieper's answer: the spherical wrist

Donald Pieper's PhD thesis at Stanford, advised by Bernard Roth, gave a compact condition:

If a 6R manipulator has three consecutive joint axes that intersect at a single point, or three consecutive axes that are mutually parallel, then the IK reduces to a polynomial of degree at most 4, solvable in closed form by the quartic formula.

The quartic formula has been known since the 16th century, so closed form means the solution computes in microseconds. In effect the thesis handed the industry a design constraint: to get real-time IK, the last three axes should intersect at a point, or be parallel. The intersecting case is the spherical wrist, and it brings a second benefit beyond the lower polynomial degree, which is that position and orientation decouple. The common intersection is the wrist center $W$, whose position depends only on the first three joints, since rotating the wrist cannot move it. Once $W$ is fixed, the last three joints set orientation alone, so the 6-by-6 system splits into two independent 3-by-3 sub-problems:

This is why you almost never find a 6-DOF industrial arm that violates the criterion. PUMA, ABB, KUKA, Stäubli, and Fanuc arms use a spherical wrist; UR arms use the parallel-axis case. Non-Pieper designs could not ship because their controllers could not solve the IK fast enough, so a 1968 result acted as a 50-year filter on what an industrial arm could be.

3. The seventh joint

A 6-DOF arm can reach any pose, but each pose has only finitely many joint configurations, up to about eight. In practice some put the elbow into the table, some exceed a joint limit, and some sit on a singularity where the Jacobian drops rank and the arm loses a direction of motion. You can also be forced to switch solution branches mid-trajectory, which shows up as a jerk in the motion. The usual fix is to add a joint.

Going to seven joints makes the system under-determined, so each reachable pose maps to a 1-dimensional curve of solutions rather than a finite set. You can feel this on your own arm: hold your hand fixed in position and orientation and swing your elbow, and it traces a circle without moving the hand. That freedom is the self-motion manifold, and it lets you place the elbow to avoid obstacles, joint limits, and singularities, and to move smoothly through the redundancy instead of switching branches. To keep closed-form IK with seven joints, apply Pieper's criterion twice: a spherical shoulder $(q_1,q_2,q_3)$, a revolute elbow $q_4$, and a spherical wrist $(q_5,q_6,q_7)$. This S-R-S structure is what the Franka Panda, KUKA iiwa, Mitsubishi PA-10, and single-arm ABB YuMi use, some with small offsets that technically break the strict condition.

S-R-S fixes the structure but leaves a whole family of solutions per pose. Shimizu's 2008 paper parameterizes the family by a single scalar $\psi$, the arm angle, which is the elbow's angular position on the circle it rides around the shoulder-wrist axis.

Once $\psi$ is fixed, every joint angle becomes an explicit closed-form function of $\psi$,

so choosing $\psi$ picks one solution and sweeping it traces the manifold. The paper also handles joint limits analytically: each constraint $q_i(\psi)\in[q_i^{\min},q_i^{\max}]$ becomes an inequality $R_i\sin(\psi+\varphi_i)\le K_i$ whose solution is one or a few arcs on the $\psi$-circle, and intersecting across all seven joints gives the complete set of feasible arm angles. Redundancy resolution then reduces to intersecting arcs on a 1-D circle and optimizing a secondary criterion within the feasible set.

4. Beyond S-R-S, and the takeaway

Shimizu assumes strict S-R-S, but arms like the Kinova Gen3 and Sawyer have offset wrists where the last three axes miss each other by a few millimeters, enough to break the derivation. The 2024 generalization by Elias and Wen replaces the arm angle with a generalized shoulder-elbow-wrist angle defined against an arbitrary reference vector, with no requirement that any three axes intersect, and decomposes the IK into composable Paden-Kahan-style geometric subproblems. The result is closed-form IK (or closed form plus a 1-D search) for essentially every modern 7-DOF arm, including offset-wrist ones. The conceptual change is the direction of the constraint: Pieper made the geometry simple so the IK was solvable, while Elias & Wen compose the IK from primitives so the geometry no longer has to be coincident, which lets designers fit the IK to the hardware.

If you are designing an arm and want analytic IK, make the last three axes meet at a point. If you already have an offset-wrist arm, numerical IK runs in milliseconds and is usually fast enough. Either way, the shape of most arms that look like a canonical robot arm traces back to a 1968 thesis that was effectively built into the hardware, and the newer offset-wrist designs are the first to move away from it.

References

1. D. L. Pieper. The Kinematics of Manipulators Under Computer Control. Ph.D. thesis, Stanford University, 1968. (Stanford AI Memo AIM-72). PDF
2. M. Raghavan, B. Roth. Inverse Kinematics of the General 6R Manipulator and Related Linkages. ASME J. Mech. Design, 115(3):502–508, 1993. DOI
3. V. Kreutz-Delgado, M. Long, H. Seraji. Kinematic Analysis of 7-DOF Manipulators. Int. J. Robotics Res., 11(5):469–481, 1992. DOI
4. M. Shimizu, H. Kakuya, W. K. Yoon, K. Kitagaki, K. Kosuge. Analytical Inverse Kinematic Computation for 7-DOF Redundant Manipulators With Joint Limits and Its Application to Redundancy Resolution. IEEE Transactions on Robotics, 24(5):1131–1142, 2008. DOI
5. A. Elias, J. T. Wen. Redundancy Parameterization and Inverse Kinematics of 7-DOF Revolute Manipulators. arXiv:2307.13122, 2024. arXiv