B-spline Trajectory Optimization

Updated 27 May 2026

B-spline trajectory optimization is a method that parameterizes trajectories with B-spline basis functions to transform continuous-time problems into finite-dimensional, tractable ones.
It leverages key properties such as local control, convex hull, and high continuity (C^(p-1)) to ensure smooth, dynamically feasible, and safe motion planning.
Applications span robotics, autonomous vehicles, spacecraft, and medical imaging, demonstrating improved runtime, reduced control effort, and enhanced collision avoidance.

B-spline trajectory optimization is a class of methods in motion planning and control that parameterize system trajectories using B-spline basis functions. This approach underlies a wide spectrum of algorithms for real-time, smooth, and constraint-satisfying trajectory generation in robotics, autonomous vehicles, spacecraft mission design, medical imaging, and manipulation. The key advantage of B-spline parameterization lies in its strong smoothness guarantees, compact representation, local control properties, and the ability to reduce continuous-time, infinite-dimensional optimization to finite-dimensional nonlinear (or convex) programs over control points.

1. B-spline Parameterization: Foundations and Properties

A B-spline curve of degree $p$ is defined by a sequence of control points $\{P_i\}$ and a (possibly non-uniform) knot vector $U=\{u_0,\dots,u_{n+p+1}\}$ . The trajectory is expressed as

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

where the basis functions $N_{i,p}(t)$ are constructed by the Cox–de Boor recursion. Key properties exploited by trajectory optimization methods include:

Local control: On each knot interval $[u_i,u_{i+1})$ , the curve depends only on $p+1$ adjacent control points, enabling local modification without affecting the entire trajectory (Ding et al., 2019).
Convex hull property: $x(t)$ lies within the convex hull of the relevant control points for each subinterval, and similar properties hold for the derivatives (Ding et al., 2019, Wang et al., 2022, Tang et al., 2019).
Continuity: A degree- $p$ B-spline is $C^{p-1}$ continuous, which is critical for generating dynamically feasible and smooth motions; e.g., cubic ( $\{P_i\}$ 0) splines ensure $\{P_i\}$ 1 continuity, suppressing jerk (Yang et al., 27 Feb 2026, Xue et al., 2023).

B-splines are used in various dimensionalities (1D, 2D, 3D, SE(3)), with the selection of degree, knot vector strategy (uniform/non-uniform), and boundary clamping adapted to the problem domain (Lang et al., 2023, Wang et al., 2021, Dorpmüller et al., 2023).

2. Formulations of Trajectory Optimization with B-splines

Decision Variables and Problem Structure

Trajectory optimization problems formulated via B-splines optimize the control points and, optionally, knot intervals. Letting $\{P_i\}$ 2 (and optionally knot vector variables), the generic problem takes the form: $\{P_i\}$ 3 Where $\{P_i\}$ 4 can encode smoothness (integrated acceleration, jerk), path length, task objectives, or domain-specific metrics (Ding et al., 2019, Choi et al., 2023, Wang et al., 2022, Xue et al., 2023).

Constraints via Convex Hull

Crucially, the convex hull property enables reduction of continuous-time constraints (velocity, acceleration, curvature, collision-avoidance) to finite sets of linear inequalities on the control points or their derivatives (Tang et al., 2019, Wang et al., 2022, Dorpmüller et al., 2023). For example: $\{P_i\}$ 5 with $\{P_i\}$ 6 closed-form in terms of $\{P_i\}$ 7.

Multi-stage and Receding-horizon Optimization

Practical methods often decompose optimization into stages: (i) an initial trajectory generation (e.g., via kinodynamic search or heuristic method), (ii) B-spline fitting or refinement over control points, and (iii) local or receding-horizon re-optimization as new information becomes available (Ding et al., 2019, Ding et al., 2019, Xu et al., 2022, Dorpmüller et al., 2023). Elastic refinement and QCQP formulations arise in post-processing to improve clearance, minimize higher-order costs, and guarantee dynamic feasibility (Ding et al., 2019, Ding et al., 2019).

3. Constraint Handling and Safety Guarantees

Dynamic Feasibility

Convex hull properties permit strict enforcement of dynamical bounds (e.g., on velocity, acceleration, jerk, snap) by bounding the relevant derivative control points. For example, for a quintic B-spline ( $\{P_i\}$ 8), the velocity constraint is enforced as: $\{P_i\}$ 9 (Tang et al., 2019, Ding et al., 2019). This admits efficient linear or quadratic programming for feasibility checks and projection.

Collision Avoidance

Collision avoidance is addressed through multiple approaches:

Convex hull checks via pre-inflated obstacles (Ding et al., 2019)
Swept-volume estimation with disc-type decomposition and incremental path flattening (IPF) (Choi et al., 2023)
Guide-point and distance field methods for dynamic obstacle avoidance using real-time perception (Xu et al., 2022)
Separating hyperplane formulation using B-spline convex hulls for both obstacles and inter-agent collision avoidance (Stoican et al., 2016, Wang et al., 2022)

For example, (Choi et al., 2023) introduces a disc-type SV model to tightly cover the vehicle swept volume, and IPF locally increases curvature penalty weights to actively flatten colliding trajectory segments.

Time and Reparameterization

Time allocation is handled via either fixed uniform knot intervals (efficient, but possibly conservative) or adaptive/non-uniform knots to distribute resolution according to dynamic/obstacle intensity (Dorpmüller et al., 2023, Lang et al., 2023). Iterative time adjustment, as in (Zhou et al., 2019), dynamically stretches knot intervals to guarantee all derivative constraints are met without introducing conservatism.

4. Algorithmic Methods and Solution Strategies

Graph-based Kinodynamic Search

A major branch employs discrete grid-based search with B-spline control points as trajectory states, transforming classical A* search into a kinodynamic framework. Nodes are constructed from tuples of overlapping control points, with edge validity and cost incorporating both control effort (smoothness) and dynamic feasibility, enabled by B-spline properties (Ding et al., 2019, Tang et al., 2019, Ding et al., 2019).

Gradient-based Direct Optimization

Another paradigm applies gradient-based optimization (e.g., L-BFGS, SQP/Interior Point, IPOPT, CasADi) over B-spline control points, using analytic gradients derived via B-spline basis and their derivatives for objective and constraints (Choi et al., 2023, Xu et al., 2022, Wang et al., 2022, Wang et al., 2021). Convexity (or QCQP structure) often arises due to the quadratic nature of the spline-based cost terms (Ding et al., 2019, Ding et al., 2019).

Piecewise-polynomial and Spline Subdivision

For strict feasibility and asymptotic optimality, piecewise Bézier (a class of B-spline) or locally refined B-spline subdivision schemes are used. These algorithms employ log-barrier or clamped penalty functions over convex hulls, combined with adaptive subdivision and safeguarded line-search to prevent intermediate infeasibility. First-order stationarity and global constraint satisfaction are achieved as subdivision and penalty parameters vanish (Ni et al., 2020).

Learning-based Extensions

In robot learning, B-spline parameterizations have been incorporated in movement primitive frameworks (BMPs), encoding both trajectory shape and statistical correlations. Learning objectives are formulated as probabilistic inference problems over the B-spline weights, providing compatibility with imitation learning and reinforcement learning methods (Liao et al., 2024, Yang et al., 27 Feb 2026).

5. Applications and Empirical Results

Robotics and Autonomous Vehicles

B-spline trajectory optimization is widely applied in UAV/quadrotor local and global replanning (Ding et al., 2019, Tang et al., 2019, Zhou et al., 2019, Ding et al., 2019, Ni et al., 2020), autonomous car racing and highway driving (Dorpmüller et al., 2023, Choi et al., 2023, Xue et al., 2023), and multi-jointed industrial manipulators (Wang et al., 2022, Liao et al., 2024, Yang et al., 27 Feb 2026). Empirical results indicate superior runtime (sub-100ms real-time operation), high success rates (≥90%) in cluttered scenarios, and smoothness/reduced control effort versus sampling-based and pointwise-parameterized planners (Ding et al., 2019, Choi et al., 2023, Ding et al., 2019, Xue et al., 2023, Wang et al., 2022, Yang et al., 27 Feb 2026).

Perception-Informed Planning

In vision-aided navigation, B-spline-optimized trajectories are coupled with real-time perception (e.g., 3D visual tracking, signed distance fields) for collision avoidance with moving obstacles. This hybrid mapping-optimization strategy enables low-latency updates (∼10ms per cycle) and high safety in dynamic environments (Xu et al., 2022).

Spacecraft Trajectory Design

Spline-based shaping methods (including B-splines) offer analytical satisfaction of end-point and time constraints for low-thrust rendezvous, drastically reducing optimization dimensionality and computation time while attaining comparable or improved performances in $U=\{u_0,\dots,u_{n+p+1}\}$ 0 and thrust bounds (Wu et al., 2022).

Imaging and Sensing

In MRI, BJORK leverages quadratic B-splines to compactly parameterize $U=\{u_0,\dots,u_{n+p+1}\}$ 1-space trajectories, supporting joint optimization of sampling and image reconstruction. Multi-level spline refinements, hardware constraint penalties (gradient/slew), and efficient backpropagation techniques enable improved image quality at aggressive acceleration rates (Wang et al., 2021).

6. Advanced Features and Extensions

Adaptive and Non-uniform Spline Placement

Non-uniform B-spline frameworks adapt the number and placement of control points online according to motion intensity, minimizing computational overhead without sacrificing representation fidelity (Lang et al., 2023). Greedy or lookup-table strategies relate IMU-derived motion indicators to control point insertion rates during high-dynamics, and marginalization ensures efficient memory use in sliding-window optimizers.

Elastic and “Tube”-based Optimization

Elastic Optimization (EO) infers ball/tube constraints along a candidate trajectory, then refines the B-spline control points via QCQP while strictly preserving dynamic and clearance bounds (Ding et al., 2019, Ding et al., 2019). Analogous ideas appear in mobile robot and car planners, incorporating incremental path flattening and disc-type swept volume calculations to achieve safe and compact paths even with complex geometry (Choi et al., 2023).

Theoretical Guarantees

Spline-based planners with convex hull and subdivision/line-search safeguard are able to guarantee continuous-time feasibility, completeness with respect to the initial guess’s homotopy class, and (when refined to the limit) first-order optimality for the true (semi-infinite constraint) problem, going beyond typical discretized or sample-based methods (Ni et al., 2020).

Learning over Spline Control Spaces

Recent work embeds B-spline representations into probabilistic movement primitive models, learning full Gaussian distributions over trajectory coefficients, which then enables joint modeling of smoothness, constraints, and uncertainty for both imitation and reinforcement learning (Liao et al., 2024, Yang et al., 27 Feb 2026).

References (by arXiv id):

(Ding et al., 2019, Ding et al., 2019, Zhou et al., 2019, Ni et al., 2020, Wu et al., 2022, Xu et al., 2022, Wang et al., 2022, Xue et al., 2023, Lang et al., 2023, Dorpmüller et al., 2023, Choi et al., 2023, Liao et al., 2024, Yang et al., 27 Feb 2026, Wang et al., 2021).

For comprehensive implementation details, cost function formulations, constraint enforcement strategies, and system-specific empirical metrics, consult the respective source papers.