Papers
Topics
Authors
Recent
Search
2000 character limit reached

Spline-Based Trajectory Representation

Updated 3 July 2026
  • Spline-Based Trajectory Representation is a method using piecewise polynomial curves (e.g., B-splines, Hermite, Bézier) defined by control points and knot vectors to model smooth, continuous motions.
  • It ensures high-order smoothness, local support, and constraint satisfaction (e.g., velocity, acceleration, collision avoidance) through properties like convex hull containment and derivative continuity.
  • Optimization techniques such as quadratic and nonlinear programming enable efficient real-time trajectory planning while meeting dynamic feasibility and safety constraints in various applications.

A spline-based trajectory representation models the time evolution of systems, motions, or signals using spline functions—piecewise polynomial curves parameterized by control points and knot vectors. In fields as diverse as robotics, vision, autonomous systems, optimal control, machine learning, and computational biology, spline models facilitate highly smooth, continuous, and differentiable trajectory representations. Each trajectory's smoothness, local/global support, and constraint satisfaction are dictated by the specific type and degree of spline, with B-splines, Hermite splines, and Bézier splines being predominant choices. This technical overview examines the formulation, theoretical guarantees, optimization strategies, key computational properties, and state-of-the-art practical applications of spline-based trajectory representations.

1. Mathematical Foundations of Spline-Based Trajectory Models

Splines are parameterized by control points and defined over a nondecreasing knot sequence. For a pp-th degree B-spline, the standard form is:

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

where PiP_i are control points in Rd\mathbb{R}^d (or on a manifold), and Ni,p(t)N_{i,p}(t) are the Cox–de Boor basis functions determined recursively:

Ni,0(t)={1uit<ui+1 0otherwiseN_{i,0}(t) = \begin{cases} 1 & u_i \le t < u_{i+1} \ 0 & \text{otherwise} \end{cases}

Ni,p(t)=tuiui+puiNi,p1(t)+ui+p+1tui+p+1ui+1Ni+1,p1(t)N_{i,p}(t) = \frac{t-u_i}{u_{i+p}-u_i}\,N_{i,p-1}(t) + \frac{u_{i+p+1}-t}{u_{i+p+1}-u_{i+1}}\,N_{i+1,p-1}(t)

Uniform splines use equally spaced knots. Continuity of order Cp1C^{p-1} at every knot and CC^\infty inside spans is intrinsic to the construction (Wang et al., 2022, Tang et al., 2019, Johnson et al., 2024). Common alternatives include Hermite splines—in which the curve is parameterized directly by knot values and derivatives—with similar piecewise-polynomial form but explicit derivative boundary conditions (Kondo et al., 13 Nov 2025, Tian et al., 5 Jun 2026). Bézier splines use parameterized Bernstein polynomials and control polygons, which are structurally equivalent over each interval to a B-spline of the same degree.

Key higher-order properties:

  • Differentiability: kk-th derivatives of a degree-x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)0 B-spline are themselves x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)1-degree splines with explicitly computable control points.
  • Convex hull property: At any x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)2, x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)3 lies within the convex hull of the x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)4 relevant control points. This is exploited for fast constraint satisfaction in velocity, acceleration, and for collision checking (Tang et al., 2019, Wang et al., 2022, Ni et al., 2020, Akhbari et al., 18 Feb 2026).
  • Manifold-valued splines: Composite de Casteljau or log-exp mapping allows intrinsic spline curves on Riemannian manifolds, e.g., for pose trajectories in x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)5 or x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)6 (Sommer et al., 2019, Nava-Yazdani et al., 2023, Ovrén et al., 2018, Johnson et al., 2024).

2. Theoretical Guarantees and Constraint Satisfaction

Spline-based representations account for physical, kinematic, and environmental constraints at the analytic level:

  • Continuity and boundary conditions: By design, splines enforce smoothness of the represented trajectory. For B-splines, x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)7 continuity follows directly; for Hermite and Bézier splines, boundary and derivative constraint equations are imposed at segment joins (Tian et al., 5 Jun 2026, Kondo et al., 13 Nov 2025, Tang et al., 2019).
  • Constraint reduction via convex hulls: Due to nonnegativity and partition of unity, enforcing box or linear constraints at each control point suffices to guarantee satisfaction for all x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)8 within the spline domain (Wang et al., 2022, Dorpmüller et al., 2023).
  • Collision avoidance: For static obstacles, signed distance or separation constraints are imposed at the control points via the convex hull property. For dynamic obstacles, time-dependent separating hyperplane strategies, receding-horizon distance fields, or continuous-time surrogate cost terms (e.g., barrier functions or hinge losses) are used (Wang et al., 2022, Xu et al., 2022, Ni et al., 2020, Covic et al., 28 Feb 2026, Tang et al., 2019).
  • Dynamic feasibility: Velocity, acceleration, and jerk limits can be checked at the derivative control points (or tight Bézier hulls). This guarantees global physical safety (Tang et al., 2019, Kondo et al., 13 Nov 2025, Dorpmüller et al., 2023).

3. Optimization and Computational Properties

Major optimization paradigms for spline parameter (control point and time) selection include:

Approach Main Features References
Quadratic programming (QP) Minimum snap/jerk, linear equality/inequality constraints (Burke et al., 2021)
Nonlinear program (NLP) Time allocation, obstacle constraints, complex costs (Dorpmüller et al., 2023, Ni et al., 2020, Kondo et al., 13 Nov 2025)
Graph search/Kinodynamic Discrete control-point search with feasibility checks (Tang et al., 2019)
Gradient-based Real-time unconstrained or soft-constrained optimization (Xu et al., 2022, Wang et al., 2022, Torzewski, 8 Apr 2025)
Recursive subdivision Adaptive spatial refinement with convex hull tests (Ni et al., 2020)
Analytic shaping Closed-form spline coefficients for fixed/flexible end data (Wu et al., 2022, Akhbari et al., 18 Feb 2026)
  • Computational complexity: For fixed-order splines and sparse constraints, QP-based approaches can achieve linear or near-linear scaling in the number of segments (Burke et al., 2021). Block-banded sparsity arises from local support, facilitating efficient Gauss–Newton or direct solvers (Johnson et al., 2024). Recursive analytic formulas for derivatives and Jacobians yield x(t)=i=0nPiNi,p(t)x(t) = \sum_{i=0}^n P_i\,N_{i,p}(t)9 per-step evaluation on Lie groups (Sommer et al., 2019).
  • Adaptive refinement: Subdivision/refinement is applied locally where collision or constraint margins are tight, maintaining feasibility while enabling efficient global optimization (Ni et al., 2020).

4. Trajectory Representation in Diverse Domains

Spline-based trajectory models are central in a range of applications:

  • Robotic motion planning: Multi-joint and Cartesian splines are used for minimum-time or minimum-jerk planning, with both soft and hard constraint enforcement. Advanced planners such as MIGHTY utilize Hermite quintic splines with collocation on Bernstein bases for efficient, unconstrained nonlinear optimization (Kondo et al., 13 Nov 2025, Wang et al., 2022).
  • Quadrotor and differentially flat systems: Flat outputs are parameterized as splines; minimum-snap/jerk is computed via QP/NLP techniques. Graph search over B-spline control points allows online replanning (Tang et al., 2019, Burke et al., 2021).
  • Autonomous vehicles: Time-scaling, shrinkable horizon, and sparse spline bases provide time-optimal planning with continuous-time feasibility and collision avoidance (Dorpmüller et al., 2023).
  • Dynamic scene reconstruction and neural fields: Hermite or B-spline based trajectory fields regularize spatial-temporal coherence, allow analytic kinematic calculation, and outperform implicit MLP-based fields in both spatial fidelity and motion realism (Song et al., 10 Jul 2025).
  • Manifold-valued trajectories in geometry and vision: Cumulative and composite B-splines are used for pose trajectories over PiP_i0, PiP_i1 with analytic derivatives, facilitating smooth visual-inertial fusion and state estimation (Sommer et al., 2019, Ovrén et al., 2018, Johnson et al., 2024).
  • Deep learning and sequence modeling: Latent space trajectories can be encoded by continuous spline curves, as in Spline-based Transformers, yielding compact, smooth representations and enabling interpretable, continuous manipulation in neural models (Chandran et al., 3 Apr 2025, Tian et al., 5 Jun 2026).
  • Computational biology: In Wasserstein space, spline interpolation via barycentric optimal transport (“Wasserstein Lane–Riesenfeld subdivision”) reconstructs continuous stochastic processes from population snapshot data (Banerjee et al., 2024).

5. Extensions: Time Parametrization, Hybrid and Manifold Spline Models

  • Non-uniform and adaptive knots: Time allocation and interval sizes can be treated as optimization variables, allowing knot-removal, dynamic resampling, and receding-horizon adaptation (Dorpmüller et al., 2023, Kondo et al., 13 Nov 2025, Torzewski, 8 Apr 2025).
  • Higher-dimensional and hybrid trajectories: Simultaneous position–orientation interpolation is achieved via hybrid position (B-spline) and orientation (quaternion-B-spline) approaches, with synchronization via arc-length parameterization and robust SLERP/Bézier-curve strategies (Akhbari et al., 18 Feb 2026).
  • Manifold splines: For PiP_i2, composite or cumulative B-splines perform Lie-algebraic interpolation, maintaining global consistency, analytic derivative recurrence, and preserving geometric structure (Sommer et al., 2019, Ovrén et al., 2018, Johnson et al., 2024, Nava-Yazdani et al., 2023). In Wasserstein space, barycentric averages replace linear combinations, with convergence guarantees (Banerjee et al., 2024).
  • Physical consistency and data denoising: Spline representations can be constructed (e.g., by integration of order-1 “radial” B-splines) to enforce strict kinematic consistency between position, velocity, and acceleration, even in the presence of missing or noisy data, forming the basis for robust trajectory filtering (Torzewski, 8 Apr 2025).

6. Benchmarking and Empirical Performance

Spline-based representations are empirically validated across domains. Highlights include:

Task/Domain Spline Form Performance Highlights References
Robotic motion planning Quintic Hermite/Bézier 100% success rate, smooth locally optimal paths (Kondo et al., 13 Nov 2025)
Quadrotor minimum-snap Polynomial/B-spline Real-time replanning, 75ms solve for 5–10 spans (Tang et al., 2019, Burke et al., 2021)
Manipulation policy learning Piecewise quadratic spline 40–50% reduction in forward-pass FLOPs, smoother action profiles (Tian et al., 5 Jun 2026)
Vision-based UAV navigation Cubic B-spline 80–120Hz online re-optimization, safe dynamic avoidance (Xu et al., 2022)
Structure from motion (SLAM) Cubic B-spline (PiP_i3) Faster convergence, PiP_i42× speedup over PiP_i5, analytic derivatives (Sommer et al., 2019, Ovrén et al., 2018)
Wasserstein barycentric inference B-spline subdivision State-of-the-art held-out interpolation error, moderate runtime (Banerjee et al., 2024)

In most technical benchmarks, spline models achieve higher trajectory smoothness (lower jerk/accel), stronger constraint satisfaction, and superior or equivalent task success rates versus grid-based, discretized, or black-box neural models.

Current and emerging directions include:

  • Real-time and adaptive computation: Efficient analytic recurrence, explicit Jacobians, and analytic time-allocation methods enable real-time, high-dimensional applications in robotics and SLAM (Sommer et al., 2019, Kondo et al., 13 Nov 2025, Covic et al., 28 Feb 2026).
  • Learning and uncertainty quantification: Spline parameterization in policy networks allows compact latent spaces and closed-form uncertainty propagation throughout the action sequence and downstream controllers (Tian et al., 5 Jun 2026, Chandran et al., 3 Apr 2025).
  • Constraint handling and exact feasibility: Adaptive subdivision and convex-hull constraint reduction guarantee exact collision-free, physically valid trajectories throughout the optimization (Ni et al., 2020, Kondo et al., 13 Nov 2025).
  • Manifold and distribution-valued interpolation: Generalizations to Wasserstein and Riemannian settings extend the utility to cell biology, hurricane tracking, and broader geometric time-series analysis (Nava-Yazdani et al., 2023, Banerjee et al., 2024).
  • Hybrid symbolic-neural formulations: Incorporation of spline experts into neural architectures yields models that combine smooth, controllable trajectories with data-driven flexibility.

References

Definition Search Book Streamline Icon: https://streamlinehq.com
References (19)

Topic to Video (Beta)

No one has generated a video about this topic yet.

Whiteboard

No one has generated a whiteboard explanation for this topic yet.

Follow Topic

Get notified by email when new papers are published related to Spline-Based Trajectory Representation.