B-Spline Trajectory Optimization

Updated 12 March 2026

B-spline trajectory optimization is a method that uses basis spline curves with local support, convex hull properties, and high-order continuity to produce smooth and dynamically feasible trajectories.
It converts continuous constraints into finite conditions on control points, ensuring dynamic feasibility, collision avoidance, and precise adherence to boundary and smoothness requirements.
Advanced optimization strategies—ranging from convex programming to gradient-based methods and adaptive knot refinement—enable real-time, robust performance across robotics, autonomous vehicles, aerospace, and industrial applications.

B-spline trajectory optimization is a class of mathematical and computational methods for generating parameterized, smooth, and dynamically feasible trajectories using B-spline (basis spline) curves as the underlying function space. B-splines possess strong local support, high-order continuity, and a powerful convex hull property, making them highly suitable for trajectory planning across robotics, autonomous vehicles, aerospace, medical imaging, and industrial path generation. This article details the mathematical structure, constraint handling, optimization methodologies, and representative applications of B-spline trajectory optimization, synthesizing foundational and recent advancements as documented in the arXiv literature.

1. Mathematical Structure and Properties of B-spline Trajectories

B-spline curves are defined by a degree $p$ (or order $k=p+1$ ), a sequence of non-decreasing knot values, and a set of control points. For a vector-valued trajectory $x(t)\in\mathbb{R}^d$ ,

$x(t) = \sum_{i=0}^{n} N_{i,p}(t)\;P_i,$

where $N_{i,p}(t)$ are the Cox–de Boor recursive B-spline basis functions and $P_i\in\mathbb{R}^d$ are the control points.

Key properties exploited in optimization include:

Local support: Each $N_{i,p}(t)$ is nonzero only over a limited interval, confining the influence of $P_i$ to a local region along the curve.
Convex hull property: For any fixed $t$ , $x(t)$ is a convex combination of up to $p+1$ neighboring $P_i$ . The entire curve lies within the convex hull defined by its control points, and the same holds for its derivatives, offering finite-dimensional surrogates for continuous constraints (Wang et al., 2022).
Continuity and differentiability: Degree $p$ ensures $C^{p-1}$ continuity, enabling control over position, velocity, acceleration, and higher derivatives (Xue et al., 2023).
Efficient evaluation and derivative computation: Derivatives of a B-spline are themselves B-splines of lower degree, and conversion to Bézier form is used for tighter convex hull bounds (Tang et al., 2019).

Non-uniform B-splines allow variable knot spacings, offering adaptive flexibility for regions with disparate motion dynamics (Lang et al., 2023).

2. Constraint Handling and Feasibility Guarantees

B-spline trajectory optimization leverages convex hull properties to reduce infinite continuous-time constraints to a finite set of conditions on the control points:

Boundary and smoothness constraints: Endpoint clamping via repeated knots ensures interpolation at endpoints. Necessary smoothness is guaranteed by degree selection ( $p\geq r$ to ensure $C^r$ continuity).
Dynamic feasibility: Velocity, acceleration, and higher-order derivative limits translate to linear or convex inequalities on control point differences or Bézier control points (Tang et al., 2019, Wang et al., 2022, Zhou et al., 2019).
Collision avoidance: For static obstacles, using the convex hull property, if the convex hull of each B-spline segment (or its derivatives for swept volumes) does not intersect obstacles, the entire curve is collision-free (Stoican et al., 2016, Choi et al., 2023).
Dynamic obstacle avoidance: Formulated via time-dependent separating hyperplanes, themselves parameterized as B-splines, and enforced as finite constraints thanks to convex hull properties (Wang et al., 2022).
Vehicle geometry: Swept volume constraints are captured using disc-based models and embedded in the optimization via costs or hard constraints (Choi et al., 2023).

For highly nonconvex constraints (e.g., complex swept volumes, joint obstacle-agent avoidance), mixed-integer or nonlinear programming is employed, sometimes using disjunctive programming or barrier methods (Stoican et al., 2016, Ni et al., 2020).

3. Optimization Methodologies and Algorithmic Structures

B-spline trajectory optimization problems are typically formulated as finite-dimensional nonlinear (often convex or quadratic) programs in the space of control points and sometimes knot intervals or breakpoints:

$\min_{P_i,\,U,\,T}\; \text{Cost}(P_i,\,U,\,T) \quad \text{s.t.}\quad \text{feasibility constraints},$

where $U$ is the knot vector and $T$ a total or piecewise time allocation.

Representative methodologies include:

Convex and quadratic programming: When costs (e.g., integrated squared acceleration, snap, jerk) and constraints are quadratic or affine in control points, convex formulations yield globally optimal solutions efficiently (Xue et al., 2023, Choi et al., 2023).
Gradient-based optimization: Nonlinear constraints or costs are addressed via L-BFGS, Levenberg–Marquardt, or sequential quadratic programming (SQP), typically coupled with efficient analytic gradients, control-point differentiation, and structured matrix assembly (Wang et al., 2022, Choi et al., 2023, Akhbari et al., 18 Feb 2026).
Graph-based and kinodynamic search: For real-time motion planning, the search for optimal control point sequences is recast as a graph search (A*, Dijkstra) over a discrete control-point grid, ensuring dynamic feasibility via (conservative or strict) convex hull checks at each node expansion (Tang et al., 2019, Ding et al., 2019, Ding et al., 2019).
Elastic and QCQP-based refinement: Output of discrete/greedy front-ends (RBK/EBK search) is refined by post-optimization of control points within "elastic tubes" of free space, handled as QCQPs (Ding et al., 2019, Ding et al., 2019).
Adaptive knot refinement and subdivision: Piecewise Bézier or B-spline segments are automatically subdivided in regions of high curvature or constraint activity; barrier methods guarantee constraint satisfaction upon refinement (Ni et al., 2020).
Sparse constraint Jacobians and symbolic coefficient transformations: State-of-the-art planners exploit symbolic basis transforms (derivative, product, sum) for efficient and scalable enforcement of inter-spline constraints (Dorpmüller et al., 2023).

4. Extensions and Domain-Specific Implementations

Autonomous Flight and Ground Robotics

Quadrotor and mobile robot planning frameworks exploit quintic B-spline parameterization for $C^4$ smoothness, closed-form integration of control cost, real-time elastic tube QCQP-based refinement, and rigorous dynamic/collision feasibility via the convex hull property (Zhou et al., 2019, Ding et al., 2019, Ding et al., 2019, Wang et al., 2022).
Incremental path flattening (IPF) and disc-based swept volume estimators yield tractable, robust collision evasion for ground vehicles under strict kinodynamic constraints (Choi et al., 2023).
Strict (Bézier-form) convex hulls deliver tighter bounds for dynamic constraint satisfaction in aggressive flight (Tang et al., 2019).

Non-uniform and cumulative B-splines, with adaptive control-point allocation, underpin real-time multi-sensor continuous-time estimation for LiDAR-Inertial-Camera odometry, outperforming uniform control-point models by matching trajectory modeling capacity to local motion intensity and reducing computational overhead (Lang et al., 2023).

Robotic Manipulation and Control

B-spline representations in the control point or action-chunk domain provide smooth, low-latency execution in manipulation. Asynchronous flow-matching and continuity-constrained refitting (CCR) architectures yield superior intra- and inter-chunk smoothness and jerk reduction in dynamic tasks compared to raw-action chunking or DCT-based approaches (Yang et al., 27 Feb 2026).

Industrial and Medical Systems

In tool path generation, hybrid B-spline–quaternion (for orientation) and minimum-jerk piecewise Bezier methods yield $C^3$ or higher continuity, precise temporal reparameterization, and real-time implementability on microcontrollers (Akhbari et al., 18 Feb 2026).
K-space trajectory design in MRI: Quadratic B-spline parameterization enables coarse-to-fine, multi-scale joint optimization of both trajectory and reconstruction network with direct enforcement of gradient slew rate and amplitude constraints, outperforming hand-designed and model-based non-Cartesian schemes (Wang et al., 2021).

Aerospace and Orbit Transfer

Piecewise cubic B-spline shaping facilitates analytical satisfaction of multi-point position, velocity, and timing constraints in low-thrust rendezvous, yielding globally robust initial guesses and highly accurate transfer profiles for global search via particle swarm optimization (Wu et al., 2022).

5. Computational Performance and Practical Considerations

B-spline trajectory optimizers can achieve real-time planning rates ( $\mathcal{O}(10$ – $100\,\mathrm{Hz})$ ), enabled by:

Low-dimensional optimization due to compact support and dimensionality reduction (e.g., $\sim$ 90% fewer decision variables for racing lines (Xue et al., 2023)).
Sparse and analytically constructible constraint Jacobians for large-scale interior-point solvers (Dorpmüller et al., 2023).
Efficient per-iteration complexity: graph-based or aggregated-node search ( $\sim$ 10–20 ms per update), QCQP refinement ( $\sim$ 20–30 ms), L-BFGS or Levenberg–Marquardt on analytic Jacobians ( $\lesssim$ 20 ms per window) (Ding et al., 2019, Tang et al., 2019, Lang et al., 2023).
Asynchronous control and inference splitting enabling $>1$ kHz trajectory evaluation on low-resource hardware (Akhbari et al., 18 Feb 2026).
Fast convergence in minimum-jerk, time-allocation hierarchies and knotted-bezier approaches (Akhbari et al., 18 Feb 2026).

6. Comparative Results and Domain Benchmarks

The comparative literature demonstrates that B-spline-based optimizers consistently outperform waypoint-based, primitive-based, or time-sampled approaches on key metrics:

Metric	B-spline Methods	Alternatives	Notable Results
Success Rate	92–100%	66–74%	UAV/AV trials (Zhou et al., 2019, Choi et al., 2023)
Traj. Smooth.	Lower $\max\|K\|$ , Jerk	Higher jerk, curvature	15–30% less max curvature/jerk
Comp. Time	10–100 ms	$>$ 1 s	$>$ 10\times $speedup</td> </tr> <tr> <td>Tracking Error</td> <td>sub-0.1 m (UAV/AV)</td> <td>$ \gg$0.1 m	Highway AVs, racing, manipulation

Safety is achieved even in dense, dynamic, and unknown environments, with rigorous guarantees on dynamic limits and clearance from obstacles (Tang et al., 2019, Wang et al., 2022, Dorpmüller et al., 2023).

7. Open Problems and Directions

Current challenges in B-spline trajectory optimization include:

Addressing limitations of uniform knot models via adaptive or non-uniform splines for complex environmental interaction (Lang et al., 2023).
Real-time local adaptation and integrity under high-dynamic or adversarial disturbances, addressed via receding horizon, robustification, and asynchronous refitting (Ding et al., 2019, Yang et al., 27 Feb 2026).
Scale-up to multi-agent or multi-modal planning frameworks, requiring tractable handling of large, coupled constraint sets (Choi et al., 2023, Stoican et al., 2016).
Further development of efficient, globally optimal solvers that combine subdivision, convex hull, and mixed-integer techniques for complex mission scenarios (Ni et al., 2020).

B-spline trajectory optimization, through advanced exploitation of spline-theoretic properties, continues to show broad applicability and state-of-the-art performance across autonomous vehicles, robotics, estimation, and industrial motion control (Tang et al., 2019, Ding et al., 2019, Choi et al., 2023, Wang et al., 2022, Akhbari et al., 18 Feb 2026, Wang et al., 2021).