Configuration Space (C-Space)

Why configuration space matters

Motion planning is deceptively hard. A robot arm is an extended object with elbows and wrists that sweep through space; asking "can this whole shape get from here to there without hitting anything?" seems to require reasoning about the geometry of every link at once. Configuration space collapses that difficulty. By choosing coordinates that fully describe the robot's state — its joint angles, or the pose of its base — every distinct configuration becomes a single point, and the machine's entire physical extent is baked into how obstacles look in that new space.

• Universal reduction. Whether the robot has 2 joints or 30, the planning problem always becomes "steer one point through a free region."
• Separation of concerns. Collision geometry is handled once, when the workspace obstacle is mapped to its C-obstacle; the planner never sees links again.
• Foundation of sampling planners. PRM, RRT, RRT*, and BIT* all operate directly on C-space points.
• Cross-domain. The same formalism plans self-driving cars (SE(2) with nonholonomic constraints), 6-DOF pick-and-place, and 30-DOF humanoid whole-body motion.
• Manipulability and singularities. Because joint space and C-space coincide for arms, the same coordinates expose Jacobian singularities and joint-limit boundaries.

How it works, step by step

The construction is the same for every robot; only the coordinates change.

1. Choose coordinates. Pick the minimal parameters q = (q₁, …, q_n) that fix the location of every point on the robot. For a serial arm these are the joint angles; for a free-flying rigid body they are position plus orientation.
2. Define the C-space. The set of all valid q is the configuration space C. Its dimension n is the number of degrees of freedom.
3. Map the robot. A function A(q) returns the set of workspace points the robot body occupies at configuration q.
4. Build C-obstacles. For each workspace obstacle O_i, the C-obstacle is CB_i = { q ∈ C : A(q) ∩ O_i ≠ ∅ } — every configuration that collides.
5. Take the free space. C_free = C \ (⋃ CB_i ∪ self-collision ∪ joint-limit violations). Only these configurations are physically valid.
6. Plan a path. Find a continuous curve τ: [0, 1] → C_free with τ(0) = q_start and τ(1) = q_goal. Any such curve is a guaranteed collision-free motion.

The elegance is that step 6 knows nothing about links, joints, or obstacle shapes. It only steers a point through an open set.

The Minkowski-sum picture and why 2D is already hard

For the simplest non-trivial case — a convex robot that can only translate (no rotation) among convex obstacles — the C-obstacle has a clean closed form. If the robot body is a set A and the obstacle is a set O, and we track the robot by a reference point, then the C-obstacle is the Minkowski sum of the obstacle with the reflected robot:

CB = O ⊕ (−A) = { o − a : o ∈ O, a ∈ A }

Intuitively, you "grow" every obstacle by the shape of the robot (reflected through its reference point), then shrink the robot to a single point. The point-robot can now slide freely anywhere the grown obstacles do not cover. This is exactly the trick behind "inflating" obstacles by a robot's radius in a costmap.

Now add rotation. A rigid robot in the plane has three degrees of freedom, so its C-space is not the flat 2D world — it is the 3-dimensional space SE(2) with coordinates (x, y, θ). For each fixed orientation θ, the C-obstacle is a Minkowski sum, but that footprint changes shape as θ varies: the robot's silhouette against the obstacle grows, shrinks, splits, and rejoins as it turns. Stacking all those slices along the θ axis produces a twisted three-dimensional solid. This is the headline result — even a robot that lives in a flat 2D world has a genuinely three-dimensional, curved configuration space, and this is before any joints are involved.

Dimension, topology, and the wrap-around problem

The dimension of C-space equals the degrees of freedom, but its shape is rarely a plain Euclidean box. Revolute joints and orientations are angles, and an angle of 0 is the same as 2π — so those coordinates wrap around. C-space is therefore a product of circles, tori, and rotation groups, not ℝⁿ.

Configuration space by robot type

Robot	DOF	C-space	Topology
Point in a plane	2	ℝ²	Flat plane
Single revolute joint	1	S¹	Circle
Planar 2R arm	2	T² = S¹ × S¹	Torus (donut)
Planar mobile base	3	SE(2) = ℝ² × S¹	Plane × circle
Rigid body in 3D	6	SE(3) = ℝ³ × SO(3)	Space × rotations
6R industrial arm	6	T⁶ (six-torus)	6 wrapped angles
Humanoid (whole body)	30+	SE(3) × Tⁿ	High-dimensional

The wrap-around matters for planning: on a torus the shortest path between two joint angles may go the "long way around" 0/2π, and a naive Euclidean distance in radians gives the wrong nearest neighbor. Sampling planners must use a metric that respects this — for angles, the geodesic distance min(|Δθ|, 2π − |Δθ|); for 3D orientation, distance on SO(3) via quaternions.

The curse of dimensionality

Why not just discretize C-space into a grid and run a graph search like A*? Because the number of grid cells explodes. If each of n dimensions is split into k intervals, the grid has kⁿ cells.

Grid cells vs. dimension (k = 100 divisions per axis)

DOF (n)	Cells = 100ⁿ	Feasible to enumerate?
2	10⁴	Trivial
3	10⁶	Easy
6	10¹²	Impractical
12	10²⁴	Impossible
30	10⁶⁰	Beyond atoms in Earth

A 6-DOF arm at a modest 100 divisions per joint already needs 10¹² cells — and computing which cells are C-obstacles requires a collision check at each one. This is why explicit C-space construction is only ever done for teaching examples and 2-to-3-DOF planners. Everything at 4 DOF and above uses sampling: draw random configurations, keep the ones a collision checker accepts, and connect them. PRM builds a reusable roadmap; RRT grows a tree biased toward unexplored Voronoi regions. Both treat C-obstacles implicitly — they never build them, they just ask "is this point free?"

Worked example: 2R planar arm

Consider a two-link arm with revolute joints at angles θ₁ and θ₂, link lengths L₁ = 0.30 m and L₂ = 0.25 m, both joints free to spin ±180°. Because each joint is an independent circle, the configuration space is the torus T² = S¹ × S¹ — a square from −π to π on each axis with opposite edges glued.

The forward kinematics place the end-effector at

x = L₁cos θ₁ + L₂cos(θ₁ + θ₂),   y = L₁sin θ₁ + L₂sin(θ₁ + θ₂)

Here θ₁, θ₂ are joint angles (rad), L₁, L₂ are link lengths (m), and (x, y) is the tool position in the workspace (m). Now place a small obstacle post in the workspace. The set of (θ₁, θ₂) pairs where either link sweeps across the post forms a curved band — the C-obstacle — winding across the torus. A pose that looks safe for the forearm may collide with the upper arm at the same tool point, because a single (x, y) is reachable by two configurations ("elbow-up" and "elbow-down"). Planning a wrist motion that dodges the post becomes: draw a curve on the donut from the start point to the goal point that never crosses the obstacle band. The physical arm's swing is invisible; only the point on the torus matters.

Common misconceptions and failure modes

• "C-space is just the workspace." No — the workspace is where the robot's body is (meters); C-space is the space of its parameters. They differ in dimension and units.
• "A 2D robot has a 2D C-space." Only a translating point does. A rigid body that can also rotate has 3 DOF and a 3D C-space, SE(2).
• "C-obstacles have the same shape as the obstacle." They almost never do. C-obstacles are inflated by the robot's geometry and, with rotation, become curved 3D solids.
• "One workspace point = one configuration." Inverse kinematics is generally many-to-one; a 6R arm can have up to 8 discrete solutions for one tool pose.
• "Angles are like real numbers." Ignoring wrap-around gives wrong distances and disconnected roadmaps; use geodesic metrics on S¹ and SO(3).
• "Grid + A* works everywhere." Above ~3 DOF the grid is astronomically large; sampling planners are mandatory.
• "C_free is one connected blob." It can be split into disconnected components by obstacles; a path exists only if start and goal lie in the same component.