⚡ electrical engineering

Simple Trajectory Smoothing for UAV Reference Path Planning Based on Decoupling, Spatial Modeling and Linear Programming

This paper presents a simple, simulation-validated trajectory smoothing method for UAV reference path planning that combines decoupling, spatial modeling, and linear programming based on a Dubins airplane model to generate algebraic control laws for flight-path angle and speed while optimizing roll angle.

Original authors: Mogens Plessen

Published 2026-03-24

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Mogens Plessen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a drone pilot, but instead of flying a toy, you are commanding a high-tech drone that needs to fly a very specific mission. You have a rough sketch of the route on a napkin—a series of dots (waypoints) the drone needs to visit.

The problem? If you just tell the drone to fly straight from dot A to dot B, then instantly turn 90 degrees to dot C, the drone will crash. Real drones (like real cars or airplanes) can't turn on a dime, and they can't change their speed or angle instantly. They need to "bank" (tilt) to turn, and they need time to climb or dive.

This paper presents a clever, three-step recipe to turn that rough, jagged napkin sketch into a smooth, safe, and efficient flight path that respects the drone's physical limits.

Here is how the method works, explained through simple analogies:

1. The "Unbundling" Trick (Decoupling)

Imagine trying to tie a knot while juggling three balls at once. It's hard. The author's first trick is to stop juggling and handle the balls one by one.

In drone physics, the movement is usually a messy mix of going up/down, turning left/right, and tilting. This method "unbundles" the problem:

Step A (The Elevator): First, it ignores the turning and just figures out how the drone should climb or dive to match the height of the waypoints. It calculates a simple "flight path angle" (how steep the nose is pointing).
Step B (The Steering Wheel): Once the up/down movement is settled, it focuses on the turning. It treats the drone like a car driving on a 2D map, figuring out how much to "tilt" (roll) to follow the curve.
Step C (The Gas Pedal): Finally, it decides how fast the drone should go. If the turn is sharp, the drone must slow down; if the path is straight, it can speed up.

By separating these tasks, the math becomes much simpler, almost like solving three easy puzzles instead of one giant, impossible one.

2. Drawing on a Map, Not a Clock (Spatial Modeling)

Usually, when we plan a trip, we think in time: "At 1:00 PM I am here, at 1:05 PM I am there."
This paper suggests thinking in distance instead of time. Imagine drawing a line on a map. You don't care when you pass a specific point, you just care about the shape of the line.

The Analogy: Think of a train track. The track exists whether a train is on it or not. The track is the "spatial path." The train's speed is just how fast it travels along that track.
Why it helps: By planning the shape of the track first (the spatial path), the math gets rid of many complicated variables. It turns a complex 3D physics problem into a simpler 2D geometry problem. It's like smoothing out a crumpled piece of paper by looking at its outline rather than worrying about how fast you are folding it.

3. The "Smart Compass" (Linear Programming)

Once the path is drawn on the map, the drone needs to know exactly how to tilt its wings to stay on that line without wobbling.

The authors use a mathematical tool called Linear Programming.

The Analogy: Imagine you are walking through a crowded hallway. You want to get to the other side as straight as possible, but you have to avoid bumping into people (constraints). You don't need a supercomputer to do this; you just make small, logical adjustments: "If I step left, I hit a wall. If I step right, I'm safe."
The Magic: The paper shows that for a drone, this "adjusting" can be solved by a very small, fast computer calculation (a "small linear program"). It's so fast that the drone can do it in real-time. It finds the perfect tilt angle to follow the smoothest possible curve.

Two Special "Flavors" of the Method

The paper offers two slightly different versions of this recipe:

The "Follow the Leader" Version: The drone tries to stay exactly on the center of the drawn line.
The "Stay to the Right" Version: Sometimes you don't want to be on the line; you want to stay to the side of it (maybe to avoid a tree or a building). This version adds a rule that says, "Always stay on the right side of the path." It's like a guardrail that keeps the drone in a safe zone.

What About Crazy Stunts? (Acrobatics)

What if the drone needs to do a loop-the-loop? The standard math breaks down because the drone is pointing straight up or upside down.

The Fix: The authors suggest a "mental rotation." Imagine the drone is doing a loop in a vertical circle. The math gets confused. So, the computer simply rotates its mental map 90 degrees, pretending the loop is actually a flat turn on the ground. It solves the problem using the simple method, and then rotates the answer back. It's a clever trick that lets the drone do stunts without needing a completely new, complex brain.

What About Tractors?

Interestingly, the same math works for tractors driving over bumpy hills!

The Connection: A tractor driving over a hill is like a drone flying over terrain. The "ground" for the tractor is the "sky" for the drone. The paper shows that if you swap the variables, this same "unbundling" and "map-drawing" method can help a tractor drive smoothly over uneven fields without tipping over.

The Bottom Line

This paper isn't about building a faster drone or a more powerful engine. It's about smarter math.

It takes a messy, difficult problem (how to fly a drone smoothly in 3D space) and breaks it down into three simple, logical steps:

Climb/Descend (Algebra).
Turn (A tiny, fast math puzzle).
Speed (Algebra).

The result is a flight path that is smooth, safe, and computationally cheap, allowing drones to fly complex missions without needing a supercomputer on board. It's the difference between a drone that jerks and crashes, and one that glides gracefully like a bird.

1. Problem Formulation

The paper addresses the challenge of trajectory smoothing for Unmanned Aerial Vehicles (UAVs) in 3D space.

Input: An initial, rough 3D path trajectory (a sequence of waypoints or an interpolated path) that may not be dynamically feasible for the UAV.
Goal: Generate a smoothed reference path plan that respects UAV dynamics and additional shaping constraints.
Model: The method is based on the Dubins airplane model, a simplified kinematic model that includes 3 inertial position coordinates ( $x, y, z$ ) and 3 angles (yaw $\psi$ , flight-path $\gamma$ , roll $\phi$ ). It assumes constant airspeed for the core derivation but allows for speed variation in the control law.
Constraints: The solution must account for:
- UAV dynamics (non-holonomic constraints).
- Physical limits (maximum/minimum roll, flight-path angles, and their rates).
- Path shaping constraints (e.g., staying to one side of a reference line for obstacle avoidance).

2. Methodology

The proposed solution utilizes a hierarchical decoupling approach combined with spatial modeling (transforming time-dependent dynamics to spatial-dependent dynamics along a reference trajectory). The method breaks down the control of three variables ( $\gamma, \phi, v$ ) into three distinct steps:

A. Decoupling Strategy

The dynamics are decoupled to simplify the control problem:

Vertical Channel ( $z$ ): Decoupled from the horizontal plane to control the flight-path angle ( $\gamma$ ).
Horizontal Channel ( $x, y, \psi$ ): Treated as a 2D path planning problem where the flight-path angle is treated as a known reference input.

B. Step 1: Flight-Path Angle Control ( $\gamma$ )

Approach: An algebraic control law is derived by integrating the decoupled $z$ -dynamics.
Mechanism: The desired $\gamma$ is calculated based on the vertical difference between consecutive points on the reference trajectory and the spatial sampling distance ( $D_s$ ).
Refinement: Rate and limit constraints are enforced via a saturation function, and a heuristic prediction step averages consecutive angles to improve tracking in steep vertical profiles.
Result: A purely algebraic solution; no optimization is required.

C. Step 2: Roll Angle Control ( $\phi$ )

Approach: This is the only step requiring optimization. The problem is reformulated in the Frenet-Serret (spatial) frame, reducing the state dimension from 3 (time-based) to 2 (spatial-based).
Linear Programming (LP): Two variations of a Linear Program are proposed to determine the optimal roll angle sequence $\{\phi_k\}$ ${ϕ_{k}}$ :
- LP1 (Centerline Tracking): Minimizes the lateral deviation from the reference path subject to roll rate and magnitude constraints.
- LP2 (Path Shaping): Adds a constraint forcing the trajectory to stay on a specific side of the reference (e.g., right-hand side) to avoid obstacles or shape the path.
Advantages: The LPs are small-scale, hyperparameter-free (no weighting matrices needed), and computationally efficient.

D. Step 3: Speed Control ( $v$ )

Approach: An algebraic upper bound on speed is derived based on the curvature of the smoothed path and the maximum allowable roll angle.
Mechanism: The speed is set to the minimum of the desired reference speed and the curvature-limited maximum speed ( $v_{max} = \sqrt{\rho_s g \tan(\phi_{max})}$ ).
Refinement: Similar to Step 1, rate and limit constraints are applied to ensure smooth acceleration/deceleration.

E. Extensions

Acrobatic Flight: The method is extended to maneuvers like loops by rotating the coordinate system (swapping $y$ and $z$ axes). This is an approximation valid for small angles in the new coordinate system but maintains the general model structure.
Ground Vehicles (Tractors): The same spatial modeling logic is applied to ground vehicles on 3D terrain, where the flight-path angle is determined by the terrain slope rather than being a control variable.

3. Key Contributions

Decoupled Control Architecture: The paper demonstrates that UAV trajectory smoothing can be achieved by separating vertical and horizontal dynamics, allowing for algebraic solutions for $\gamma$ and $v$ , and reserving optimization only for $\phi$ .
Spatial Modeling & Dimensionality Reduction: By transforming the problem into the spatial domain, the authors reduce the state dimension and naturally handle spatial constraints (like curvature limits) without complex time-parameterization.
Hyperparameter-Free Optimization: The proposed Linear Programs require no tuning of weighting matrices (unlike typical Model Predictive Control), making the method robust and easy to implement.
Generality: The framework is shown to be applicable not only to fixed-wing UAVs but also adaptable for quadrotors (via the 6-variable model) and ground vehicles on 3D terrain.

4. Results

Simulations: The method was validated through numerical experiments (visualized in Figures 2–5 of the paper).
Performance:
- Successfully smoothed "edgy" reference paths with sharp turns into feasible trajectories.
- Demonstrated the ability to track reference paths while adhering to strict roll and flight-path limits.
- LP2 successfully enforced path shaping constraints (keeping the UAV to one side of the reference).
- The method handled varying reference speeds and complex 3D waypoints effectively.
Computational Efficiency: The small size of the LPs ensures fast solution times, suitable for real-time or high-level mission planning.

5. Significance and Limitations

Significance:

Simplicity: The method offers a "simple" alternative to complex non-linear optimization or heavy MPC approaches for high-level path planning.
Scalability: The low computational cost makes it suitable for swarm planning or formation flying where many trajectories must be generated quickly.
Bridging Gaps: It fills a research gap by exploiting the specific structure of the Dubins airplane model for control decoupling, a feature previously underutilized in 3D path planning literature.

Limitations:

Model Simplification: The Dubins model ignores aerodynamics, wind, sideslip, and low-level control delays. It is intended for high-level reference planning, not low-level tracking.
Singularity Issues: The spatial modeling approach is sensitive to singularities if the reference path has instantaneous 180° turns or extreme curvature changes. The paper suggests a hybrid approach (time-domain linearization for the initial step) to handle such extreme cases.
Acrobatic Approximation: The extension for acrobatic flight involves a model approximation valid only for specific coordinate orientations and small angles in the transformed frame.

In conclusion, the paper presents a robust, computationally efficient, and mathematically elegant framework for generating smooth, dynamically feasible 3D reference paths for UAVs, leveraging decoupling and spatial linear programming to avoid the complexity of full non-linear optimization.