Nonholonomic Constraints

Nonholonomic constraints restrict allowable velocities at each configuration, not the configuration itself. A differential-drive robot cannot slide sideways instantaneously—its velocity vector must align with its heading. However, it can reach any (x, y, θ) pose through sequential maneuvers like parallel parking. This distinguishes nonholonomic constraints from holonomic ones, which reduce the dimensionality of reachable space.

The canonical nonholonomic constraint is the rolling-without-slipping condition for wheels. For a wheel of radius r with angular velocity ω, the constraint is ẋ = rω cos θ and ẏ = rω sin θ. This Pfaffian constraint takes the form A(q)q̇ = 0 and is non-integrable—no function f(q) = 0 exists whose derivative yields the constraint. This non-integrability is the mathematical hallmark of nonholonomic systems.

Nonholonomic systems can generate motion in constrained directions through cyclic sequences of admissible motions. The Lie bracket of two control vector fields quantifies this effect. For a car-like robot, the sequence: steer left → drive forward → steer right → reverse → steer straight produces a net lateral displacement. This principle underlies sinusoidal steering strategies and proves that nonholonomic constraints do not reduce the accessible configuration space.

A nonholonomic system is small-time locally controllable (STLC) if it can reach any point in an arbitrarily small neighborhood in arbitrarily small time. This is verified through Chow's theorem: if the Lie algebra generated by control vector fields spans the full tangent space at every configuration, the system is controllable. This guarantees that path planners can find solutions despite instantaneous motion restrictions.

Nonholonomic constraints appear across mobile robotics:

• Differential-drive robots: cannot move perpendicular to wheel axis
• Car-like (Ackermann) vehicles: minimum turning radius limits curvature
• Unicycle model: single wheel with forward velocity and steering rate
• Trailer systems: each additional trailer adds nonholonomic constraints
• Snake robots: body segments constrained by lateral friction anisotropy Each requires specialized path planning algorithms that respect these velocity-level restrictions.

Nonholonomic constraints force path planners to consider kinematic feasibility beyond collision avoidance. Standard RRT and PRM algorithms must be extended to steer functions that generate admissible curves—such as Dubins paths (forward-only, bounded curvature) or Reeds-Shepp curves (forward and reverse). Without these, planned paths may be geometrically valid but physically impossible to execute, requiring post-processing smoothing or constrained optimization.

The standard mathematical form for expressing nonholonomic constraints as a linear relationship between generalized velocities: A(q)q̇ = 0. This matrix equation captures the fact that admissible velocities must lie in the null space of A(q) at any configuration q. The rolling-without-slipping condition for a unicycle or car-like robot is canonically expressed in this form, and the constraint is nonholonomic precisely because A(q) is not an exact differential—it cannot be integrated to yield an equivalent position-level constraint.

The definitive mathematical test for determining whether a set of velocity constraints is holonomic or nonholonomic. The theorem states that a distribution of admissible velocities is completely integrable if and only if it is involutive—meaning the Lie bracket of any two vector fields spanning the distribution remains within the distribution. For wheeled mobile robots, the Lie bracket of the forward and steering vector fields typically generates a sideways motion direction, proving the distribution is not involutive and the constraint is genuinely nonholonomic.

A counterintuitive consequence of nonholonomic constraints: even though a robot cannot move directly in a constrained direction, it can achieve net displacement in that direction through cyclic, out-of-phase motions. This is the mechanism behind parallel parking—alternating forward steering and backward steering generates a net lateral shift. Mathematically, the Lie bracket [g₁, g₂] of two control vector fields produces motion along a new direction not directly controllable, enabling full configuration space reachability despite velocity-level restrictions.

A canonical state-space representation into which many nonholonomic systems—including car-like robots and tractor-trailer configurations—can be transformed via feedback and coordinate change. The chained form has a triangular structure where each state's derivative depends only on the preceding state and the control inputs. This normalized form simplifies controller design: path-following and stabilization controllers designed for the chained form apply to any system transformable into it, making it a foundational tool in nonholonomic control theory.

A fundamental limitation established by Roger Brockett in 1983: no continuous, time-invariant, pure-state feedback law can asymptotically stabilize a nonholonomic system to a point. This explains why smooth proportional controllers fail for parking problems. Acceptable solutions must violate at least one condition—using discontinuous feedback, time-varying control (e.g., sinusoidal steering), or hybrid strategies. This theorem directly motivates the use of model predictive control and trajectory-based approaches rather than naive setpoint regulation for wheeled robots.

A structural property where all system states and control inputs can be expressed as algebraic functions of a carefully chosen set of flat outputs and their derivatives. For a car-like robot, the rear-axle midpoint position serves as a flat output—the steering angle and velocity are recoverable from its time derivatives. Flatness transforms trajectory generation from a constrained differential problem into a simpler curve-smoothing problem in the flat output space, guaranteeing that any sufficiently smooth path in the plane corresponds to a dynamically feasible trajectory.