Nonholonomic Constraint

The defining mathematical property of a nonholonomic constraint is that it is non-integrable. This means the velocity-level restriction cannot be converted into an equivalent position-level constraint. Formally, a constraint of the form A(q)q̇ = 0 is nonholonomic if it cannot be integrated to a function Φ(q) = constant. This results in a system where the set of reachable positions is larger than the set of positions that satisfy the velocity constraint if it were integrable.

Nonholonomic systems can reach any configuration in their configuration space but require specific, often non-trivial, paths to do so. This is a direct consequence of non-integrability. For example, a car parked in a tight spot can reach any position and orientation, but it must execute maneuvers like parallel parking—it cannot simply slide sideways. The system's mobility is not defined by a reduced configuration space but by the Lie bracket of its allowable motion directions, which generates new, indirect directions of movement.

Nonholonomic constraints typically lead to underactuation, where the number of independent control inputs is less than the number of degrees of freedom to be controlled. A car has three configuration variables (x, y, θ) but only two control inputs (steering angle and forward/backward drive). This mismatch prevents direct, independent control of all configuration variables simultaneously, necessitating complex planning and feedback control strategies like nonlinear control or model predictive control (MPC).

These constraints are ubiquitous in wheeled and rolling systems:

• Car-like vehicles (Ackermann steering): Cannot move directly sideways.
• Bicycles and motorcycles: Similar kinematic restrictions to cars.
• Differential-drive robots: Two independently driven wheels on a common axis; the robot's instantaneous motion is always tangent to its orientation.
• Rolling without slipping contact: A rolling wheel or ball's point of contact has zero instantaneous velocity, coupling rotational and translational motion.
• Spacecraft with thrusters: Conservation of angular momentum acts as a nonholonomic constraint on attitude control.

Nonholonomic constraints are expressed as Pfaffian constraints: A(q)q̇ = 0, where q is the vector of generalized coordinates, q̇ is the velocity, and A(q) is a constraint matrix. The constraint is nonholonomic if the distribution defined by A(q) is not involutive (i.e., not closed under the Lie bracket operation). Testing for this involves calculating the accessibility Lie algebra from the system's vector fields. The famous Frobenius theorem provides the condition for integrability (holonomy).

Planning for nonholonomic systems is significantly harder than for holonomic systems. Key algorithmic considerations include:

• Sampling-based planners (RRT, PRM) must use specialized steering functions or local planners that respect the constraints to connect sampled states.
• Trajectory optimization must include the nonholonomic equations as hard constraints in the nonlinear programming (NLP) problem.
• The sub-Riemannian geometry of the system dictates the structure of optimal paths, which are often sequences of arcs and straight lines (e.g., Reeds-Shepp curves for cars).
• Pure graph search algorithms (A*, Dijkstra) on a grid are often insufficient unless the state space includes orientation.

A holonomic constraint is a kinematic restriction that can be expressed as an equation involving only the system's configuration variables (positions) and time, without their derivatives. This means it limits the set of reachable configurations directly. For example, a robot arm's joint limits or the fixed length of a pendulum rod are holonomic constraints. The key distinction from a nonholonomic constraint is integrability: holonomic constraints are integrable, meaning they reduce the degrees of freedom of the configuration space itself.

An underactuated system is a dynamical system that has fewer independent control inputs than degrees of freedom. Nonholonomic constraints often lead to underactuation. For instance, a car (with steering and throttle) has two controls but three configuration variables (x, y, heading θ), making it underactuated. This mismatch means the system cannot instantaneously accelerate in any arbitrary direction in its configuration space, which is precisely what nonholonomic constraints describe. Planning for underactuated systems requires complex maneuvers like parallel parking.

A Pfaffian constraint is a linear constraint on the generalized velocities of a system, expressed in the form A(q)q̇ = 0, where q is the configuration and A(q) is a constraint matrix. Nonholonomic constraints are a specific type of non-integrable Pfaffian constraint. This formulation is the standard mathematical starting point for analyzing nonholonomic systems. Determining whether a Pfaffian constraint is holonomic (integrable) or nonholonomic is done using tools from differential geometry, such as Frobenius' theorem.

The Lie bracket, denoted [X, Y], is an operation on vector fields that measures the failure of their flows to commute. In nonholonomic systems, it is central to analyzing controllability. While instantaneous motion may be restricted to directions defined by constraint-permissible vector fields, applying sequences of allowed motions (like forward, turn, backward for a car) can generate new, effective directions of motion. This is captured mathematically by the Lie bracket. If the Lie algebra generated by the permissible vector fields spans the entire tangent space, the system is controllable despite its constraints.

Configuration Space (C-Space) is the space of all possible configurations (positions and orientations) of a robot. A holonomic constraint reduces the dimensionality of the C-space (e.g., a pendulum's C-space is a circle, not a plane). A nonholonomic constraint does not reduce the dimensionality of the C-space; instead, it restricts the allowable tangent vectors (velocities) at each point. This means the robot can theoretically reach any point in its C-space, but only via specific, often non-straight, paths. Motion planning occurs within this space.

The Reeds-Shepp car is a canonical mathematical model for a car-like robot with a nonholonomic constraint. It assumes a car that can move forward and backward, with a minimum turning radius. The Reeds-Shepp algorithm provides an optimal (shortest) path between any two configurations for this model, composed of sequences of straight segments and arcs of the minimum turning circle. It is a foundational solution in nonholonomic motion planning and is often used as a heuristic or local planner within broader sampling-based algorithms.