The Shoulder-Elbow-Wrist Trick
Hold a glass of water out in front of you and keep your hand perfectly still. Now swing your elbow up and down. The glass does not move — yet your arm is reconfiguring the whole time. That is possible because your arm has more joints than the task strictly requires: placing a hand at one pose (position plus orientation) in space needs only six numbers, but your shoulder, elbow, and wrist together offer seven turning degrees of freedom. The spare one is freedom you can spend however you like.
Robot arms can have this same gift. The earlier rungs of this ladder built up the core machinery: forward kinematics (joint angles in, hand pose out), inverse kinematics (target pose in, joint angles out), and the Jacobian that links joint speeds to hand speed. This guide is about what happens when there are extra joints to go around — and how that extra room becomes a tool rather than a nuisance.
Redundancy: One Hand Pose, Many Arm Shapes
An arm is said to have kinematic redundancy when it has more joint degrees of freedom than the task in front of it requires. A seven-joint arm asked to put its hand at a six-degree-of-freedom pose has one degree of redundancy; the same arm asked only to touch a point (ignoring orientation) has even more. Redundancy is relative to the task, not a fixed property of the hardware.
The consequence is profound. For a non-redundant arm, a reachable hand pose usually corresponds to a small, finite set of joint solutions (the classic 'elbow-up versus elbow-down' choices). For a redundant arm, a single hand pose corresponds to a whole continuous family of joint configurations — infinitely many ways to be in the same place. The inverse-kinematics problem stops having one answer and starts having a curve, surface, or volume of answers in joint space.
Choosing wisely among that family is its own discipline, called redundancy resolution. The idea is to keep the hand exactly where the task wants it while spending the leftover freedom on a secondary goal — staying away from joint limits, keeping elbows clear of an obstacle, or staying nimble. The hand pose is the hard constraint; everything else is a preference you optimize underneath it.
Null-Space Motion: Moving Without Going Anywhere
How does a robot swing its elbow without moving its hand? The answer lives in the Jacobian. Recall that the Jacobian turns joint velocities into hand velocity. For a redundant arm, there exist whole combinations of joint velocities that the Jacobian maps to zero hand velocity — the joints move, but their effects cancel out at the hand. That set of self-cancelling joint motions is called the null space of the Jacobian, and moving along it is null-space motion.
This is exactly the glass-of-water trick, made precise. Null-space motion is internal reconfiguration: the arm rearranges itself while the end-effector stays frozen on its target. Because it costs the task nothing, it is the perfect place to pursue those secondary goals — dodging an obstacle that drifts toward the elbow, steering joints away from their mechanical limits, or backing away from a kinematic singularity where the arm would otherwise lose the ability to move in some direction.
Joint command = primary task + null-space term
q_dot = J_pseudoinverse * x_dot_desired <- moves the hand
+ N * z <- moves the elbow only
where N projects any desired joint motion z
into the null space, so it changes the arm's
shape WITHOUT changing the hand's velocity.You can even stack priorities. The hand pose can be the top priority, obstacle avoidance the next, and a comfortable posture the last — each lower goal allowed to act only within the freedom the higher ones leave behind. This layering of tasks in the null space is the seed of the whole-body controllers used on modern humanoids.
Manipulability: A Map of How Easily the Hand Can Move
Even when an arm can reach a point, it is not equally agile in every direction there. Near a stretched-out, almost-straight pose, the hand glides easily sideways but resists moving along the line of the arm — like your own arm when fully extended. The manipulability ellipsoid captures this at a glance: imagine asking the joints to move at unit effort in all directions and plotting how fast the hand responds. The result is an ellipsoid, fat along easy directions and pinched along hard ones.
A round, plump ellipsoid means the hand is nimble in every direction — a dexterous pose. A flat, pancake-thin ellipsoid is a warning: the arm is near a singularity, and along the squashed axis it can barely move at all. Engineers boil the ellipsoid down to a single manipulability number (loosely, its volume) and use it as a quality score. Redundancy resolution can then spend null-space motion specifically to keep that score high, steering the arm toward poses where it stays quick and responsive.
This sharpens an idea from earlier guides. The full set of points a hand can touch is the reachable workspace; the smaller region where it can also arrive in any orientation it likes is the dexterous workspace. Manipulability adds a third layer of nuance: not just whether a pose is reachable, but how gracefully the arm can move once it gets there.
The Frontier: Whole-Body Control and Learned Kinematics
Push redundancy to its limit and you arrive at the humanoid. A humanoid robot has dozens of joints across two arms, a torso, and two legs, so almost every task is wildly redundant. The modern answer is whole-body control: treat the entire body as one big redundant system and solve, at every instant, for joint motions that honor a stack of prioritized goals — keep balance first, place the hands second, stay comfortable third — exactly the null-space layering from before, scaled up to the whole machine.
Classical kinematics is also gaining a learning-based partner. Numerical inverse kinematics can be slow and can get stuck; increasingly, a neural network is trained to propose good joint configurations in one shot, which a classical solver then polishes. And at the far edge, vision-language-action models and imitation learning let a robot map camera images straight to motion, absorbing redundancy resolution implicitly from human demonstrations rather than from hand-written rules.