Planning Kinematically Feasible Paths for Nonholonomic Robots

337 concepts264 papers1 proposal

When we think about path planning geometrically, we tend to treat a robot as a dimensionless point moving freely across a map. If there is an open space, the robot can move into it; if it needs to change direction, it turns instantly. This assumption of unconstrained, omnidirectional movement is known as a holonomic assumption.

But real physical systems rarely work this way. A car cannot slide sideways into a parallel parking spot; it must execute a series of forward and backward maneuvers dictated by its steering geometry. A fixed-wing drone cannot stop and hover; it must maintain a minimum forward velocity and bank along a wide arc to change its heading. These physical limitations are nonholonomic constraints—velocity relationships that cannot be integrated into purely positional constraints.

When path planners ignore these constraints, they generate trajectories that are geometrically valid but physically impossible for the hardware to execute. Bridging this gap requires planning algorithms to embed the kinematics of the vehicle directly into the optimization or search process.


The Cost of the Holonomic Illusion

As Charles E. Thorpe argued in Path Relaxation: Path Planning for a Mobile Robot, standard grid searches and optimization routines lack the mechanisms to account for a vehicle's minimum turning radius. When a holonomic planner encounters a sharp corner, it produces a jagged, instantaneous turn. For a real car-like or differential-drive robot, executing this path requires coming to a complete stop, rotating, or violating its physical steering limits.

To generate kinematically feasible paths, a navigation system must ensure that every segment of a planned route respects the vehicle's mechanical boundaries.


Strategy 1: Embedding Kinematics into the Search Space

The most direct way to solve the nonholonomic problem is to change what the planner is searching. Instead of searching a simple 2D grid of coordinates $(x, y)$, the planner must search a higher-dimensional configuration space that includes orientation $(\theta)$ and velocity $(v)$.

Early autonomous systems achieved this through geometric constructions. In The Navigation System of the JPL Robot, Alan M. Thompson incorporated vehicle width and minimum turning radii directly into the path generation phase using tangent circles. By constructing paths out of circular arcs and straight lines, the planner guaranteed that the robot's physical dimensions and mechanical limits were satisfied by construction.

Modern graph-based planners scale this intuition using state lattices. Instead of connecting neighboring pixels in a grid, a state lattice planner connects states using precomputed, kinematically feasible motion primitives.

Holonomic Grid Search (Infeasible)       State Lattice Search (Feasible)
       [ ] --- [ ] --- [ ]                      [ ] \
        |       |       |                            \  --- [ ]
       [ ] --- [ ] --- [ ]                      [ ] --- /
        |       |       |                      (Transitions match vehicle steering)

Yoon et al. leverage this approach in Recursive Path Planning Using Reduced States for Car-Like Vehicles on Grid Maps. By restricting the search graph to transitions that a car-like steering system can actually perform, their algorithm can navigate highly constrained environments—such as narrow corridors requiring multi-point turns—with the certainty that the resulting path is immediately executable by the hardware.


Strategy 2: Decoupling and Refinement

Searching a high-dimensional space $(x, y, \theta)$ is computationally expensive because the graph grows exponentially with each added dimension. To maintain real-time performance, a powerful alternative strategy is to decouple global pathfinding from local kinematic feasibility.

Jean-Claude Latombe pioneered this approach in A Fast Path Planner for a Car-Like Indoor Mobile Robot. Rather than computing a complex nonholonomic path from scratch, the planner works in two distinct phases:

  1. Global Holonomic Draft: A fast, low-dimensional planner finds a collision-free path ignoring all steering constraints. This establishes the topological route around obstacles.
  2. Kinematic Refinement: A local planner transforms this unfeasible geometric path into a sequence of valid nonholonomic segments (such as Dubins curves or Reeds-Shepp paths).

This decoupling provides a massive computational advantage. The heavy lifting of obstacle avoidance is handled in a simple space, while the complex kinematics are resolved locally, accelerating planning speeds by orders of magnitude.


Strategy 3: Exploiting Mathematical Flatness

For highly agile or dynamic systems, searching graphs or smoothing paths can still be too slow. Can we instead use algebra to guarantee kinematic feasibility?

One approach is to transform the system's non-integrable velocity equations into a highly steerable canonical structure. Richard M. Murray demonstrated this in Nonholonomic motion planning: steering using sinusoids, showing that a broad class of nonholonomic systems can be converted into "chained systems." By applying sinusoidal control inputs at integrally related frequencies, the system can be steered precisely to a target configuration, turning a difficult differential equation problem into a straightforward algebraic calculation.

A closely related mathematical property widely used today is differential flatness. In a differentially flat system, all nominal states and control inputs can be mapped directly to a carefully chosen set of "flat outputs" and their derivatives.

Morando et al. apply this concept to fixed-wing aircraft in Trajectory Planning and Control for Differentially Flat Fixed-Wing Aerial Systems. For a fixed-wing UAV, generating trajectories that respect aerodynamic lift, forward velocity, and banking constraints is incredibly complex. However, because the system is differentially flat, the authors can plan a smooth, unconstrained trajectory in the flat output space using algebraic curves like Bernstein polynomials.

Because the mathematical mapping inherently accounts for the vehicle's physics, transforming this flat trajectory back into the real world guarantees that the aircraft's nonholonomic constraints are perfectly satisfied. This allows kilometers-long, flyable trajectories to be calculated in milliseconds.

Key Foundational Work

The Navigation System of the JPL Robot (1977) — contributed 3 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 253 papers
  2. Version 1 · 2026-10-04 · 355 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_non_holonomic_motion_constraints_d7ae0dc5,
  title = {non-holonomic motion constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-holonomic-motion-constraints/rdr_d7ae0dc57d9947918c330f02f0eb2a25},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 2, frozen text):

@misc{lacuna_non_holonomic_motion_constraints_d7ae0dc5_v2,
  title = {non-holonomic motion constraints},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/non-holonomic-motion-constraints/rdr_d7ae0dc57d9947918c330f02f0eb2a25/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake