Decoupled Path-Speed Trajectory Planning

• The decoupled path-speed approach separates the trajectory generation into path planning and speed planning, enabling collision-free and dynamically feasible motion.
• It leverages convex and dynamic programming techniques to reduce problem dimensionality, resulting in computational efficiency and real-time performance.
• The method is widely applied in autonomous vehicles and robotics, delivering measurable improvements in safety, smoothness, and operational speed.

The decoupled path-speed approach is a foundational paradigm in trajectory planning for autonomous vehicles, mobile robots, and robotic manipulators. It separates the joint problem of generating a collision-free, dynamically feasible, and comfortable state–space trajectory into two stages: (1) path planning—a geometric or spatial optimization typically subject to kinematic and static obstacles—and (2) speed planning, where given the fixed path, a time-parameterized speed profile is synthesized under dynamical, actuation, and comfort constraints. Decoupling reduces the overall problem's dimensionality, enabling algorithmic efficiency and real-time feasibility, and underlies a wide range of frameworks deployed in both research and industrial systems (Tariq et al., 6 May 2025, Kapania et al., 2019, Li et al., 2021, Marauli et al., 23 Oct 2025, Consolini et al., 28 Oct 2025, Zhou et al., 2020, Consolini et al., 2021).

1. Core Principles and Computational Rationale

Decoupling transforms the generally intractable full space–time optimal control problem into two subproblems amenable to fast optimization algorithms. The first, path planning, operates in a geometric or Frenet frame, searching for a collision-free, smooth, and kinematically feasible curve. The second, speed planning, treats the geometric path as fixed and computes the optimal speed profile, typically minimizing a weighted sum of travel time, energy, jerk, or deviation from a reference speed, subject to actuation and safety constraints. The reduction in search-space dimension—from typically 3–5D to 1–2D per stage—enables convexity or pseudo-convexity, faster convergence, and tractable online replanning (Tariq et al., 6 May 2025, Li et al., 2021, Zhou et al., 2020).

In path planning, non-holonomic constraints (such as curvature limits) and collision avoidance can be explicitly enforced with efficient spatial models, e.g., space-domain bicycle models (Tariq et al., 6 May 2025) or convex feasible bubbles (Zhou et al., 2020). Speed planning then leverages the fixed geometric context to employ convex formulations in the arc-length or path-parameter domain, allowing for linear/quadratic programming and dynamic programming solutions (Marauli et al., 23 Oct 2025, Consolini et al., 28 Oct 2025, Consolini et al., 2021). This structural separation also naturally supports modular, maintainable implementation in embedded and robotics systems.

2. Formal Methodologies and Representative Algorithms

Several precise algorithmic templates instantiate the path-speed decomposition, varying in representation, constraints, and application domain:

Frenet Corridor Planner (FCP) (Tariq et al., 6 May 2025): Path planning is performed in the Frenet (lateral–longitudinal) frame, with obstacle representation via safety-augmented bounding boxes and convex hulls, generating a drivable corridor. The path optimizer solves a nonlinear program for laterally optimal offset and steering, with objectives reflecting smoothness, boundary clearance, and dynamic-obstacle risk. Kinematics are enforced with a discretized, space-domain bicycle model. The optimized path $\{s_k, d_k\}$ is then passed to a real-time speed planner (e.g., multi-profile QP) that minimizes jerk and matches a reference speed, with constraints including collision-avoidance in time-parameterized zones.

Iterative Two-Step Racing Trajectory (Kapania et al., 2019): The method alternates between (a) computing the minimum-time longitudinal speed profile for a fixed path curvature via forward-backward integration (ensuring tire friction, engine, and braking limits), and (b) updating the steering path by convex curvature minimization, enforcing both vehicle dynamics and track boundaries. Convergence yields a racing-optimal (but not globally optimal) trajectory.

DL-IAPS + PJSO (Zhou et al., 2020): Dual-loop Iterative Anchoring Path Smoothing (DL-IAPS) employs sequential convex programming with polygonal “bubbles” for exact collision avoidance and curvature limits. Piece-wise Jerk Speed Optimization (PJSO) solves a convex QP for the time-parameterization, enforcing speed, acceleration, jerk, and curvature-induced lateral acceleration constraints.

Bezier Polynomial Speed Planning with Trapezoidal Corridors (Li et al., 2021): A path-level corridor defines feasible regions in the time–station (S–T) graph derived from dynamic obstacle motion. Dynamic programming generates a way-pointed safe corridor, within which the time-profile $s(t)$ is parameterized by piecewise Bezier polynomials and optimized via convex QP, guaranteeing provable safety under affine-in-time corridor bounds.

Parameter-Linear Formulation for Manipulator Path Following (Marauli et al., 23 Oct 2025): Fixing a geometric path $q(s)$, a linear program over path-speed squared variables $z_k$ directly optimizes performance (e.g., average path-speed) under kinematic and torque constraints, with all path-dependent quantities precomputed offline and only the 1D speed profile optimized online.

Convex and DP-Based Speed Planners (Consolini et al., 28 Oct 2025, Consolini et al., 2021): The speed planning stage on a predefined path is formulated as a convex program (LP, QP, or SOCP), or as a dynamic program over a small candidate set. Nonconvexities arising from energy minimization, power limits, or jerk are managed either by convex relaxation (guaranteed exact under mild conditions) or by specialized sequential algorithms that alternate between convex approximations and line-search optimization.

Stage	Model/Technique	Typical Constraints
Path planning	Frenet, Cartesian, graph	Curvature, geometry, obstacles
Speed planning	QP, DP, LP, Bezier-Bernstein	Speed/acceleration/jerk, time energy, collision zones

3. Theoretical Guarantees and Performance

The decoupled approach admits several theoretical properties and computational metrics:

• Predictable real-time performance: Representative implementations solve both subproblems within 0.03–0.07 s on a single CPU core, supporting 10 Hz planning loops in realistic autonomous driving and manipulation scenarios (Tariq et al., 6 May 2025, Zhou et al., 2020, Li et al., 2021).
• Convexity and global/efficient optima: Formulations leveraging the convex-hull property of Bezier curves, or boxed interval dynamics via monotone maps, ensure that feasibility and optimality are tractable and exactly enforced where proven (Marauli et al., 23 Oct 2025, Li et al., 2021, Consolini et al., 28 Oct 2025).
• Provable feasibility and safety: Algorithms guarantee no-collision motion in the presence of moving obstacles and respect hard physical constraints (e.g., curvature-induced speed limits, acceleration bounds).
• Performance in benchmarks: On complex obstacle scenarios in simulation and hardware, decoupled planners show improvements of 15–20% in minimum safety distance and order-of-magnitude reductions in heading change (smoother trajectories) compared to graph-based or sampling baselines (Tariq et al., 6 May 2025). Racing implementations match or approach expert human drivers, with trajectory tracking errors on the order of centimeters and angular deviations below 1.5° (Kapania et al., 2019, Zhou et al., 2020).

4. Limitations and Practical Implications

Despite advantages in computational efficiency and modularity, decoupled methods exhibit inherent trade-offs:

• Conservatism in space–time coupling: The fixed-path approach cannot exploit path adjustments to compensate for timing constraints induced by dynamic obstacles, sometimes yielding conservative or suboptimal trajectories (e.g., in tightly coupled overtaking or aggressive merge scenarios) (Tariq et al., 6 May 2025, Zhou et al., 2020).
• Sensitivity to geometric infeasibility: If the path stage returns an over-constrained or suboptimally feasible path, the speed planner may be unable to find a travel time- or energy-optimal profile.
• Linearization and model-validity constraints: Some algorithms assume small-angle dynamics, precise curvature approximation, or local convexity. Significant deviations (e.g., high curvature, extreme friction conditions) may require adaptive tuning or re-computation (Kapania et al., 2019, Consolini et al., 28 Oct 2025).
• Inability to enforce concurrent spatio-temporal interactions:

Multi-agent scenarios or tightly coordinated maneuvers demand either rapid re-planning or partial recoupling approaches (Li et al., 2021, Zhou et al., 2020).

This suggests that applications with tightly coupled timing-path interdependencies may demand extensions or hybridizations with coupled planners, at the cost of computational overhead.

5. Algorithmic Innovations and Recent Trends

Recent research advances multiple directions within the decoupled path-speed paradigm:

• Convexification and lattice-based feasibility: Novel convex reformulations for speed planning achieve exactness under provable conditions, enabling the use of efficient dynamic programming and providing guarantees of feasibility and optimality (Consolini et al., 28 Oct 2025).
• Structured representations: Use of Bezier polynomials with trapezoidal corridor invariants tightly exploit the convex-hull property to guarantee collision-free trajectories with real-time solvers (Li et al., 2021).
• Sequential convex approximation: Line-search and sequential convex programming methods efficiently solve nonconvex minimum-time with jerk (comfort) constraints (Consolini et al., 2021), with per-iteration complexity managed via banded/tridiagonal KKT systems and convergence to KKT stationarity.
• Parameter-linear and B-spline speed optimization for manipulators: The explicit path-wise LP/B-spline design avoids singularities and reduces online computation by over 50% compared to classical time-minimization approaches (Marauli et al., 23 Oct 2025).
• Empirical system validation: Integrated deployments in systems such as Baidu Apollo and on scale-model platforms demonstrate real-time, robust, and precise operation in cluttered and uncertain environments (Zhou et al., 2020, Tariq et al., 6 May 2025).

6. Application Domains and Future Directions

The decoupled path-speed approach is foundational in autonomous driving, industrial AGVs, mobile robots in warehouse or factory logistics, and articulated manipulator time-optimal and energy-optimal path following. Its flexibility for varying environments, precise constraint handling, and compatibility with real-time control pipelines make it broadly applicable.

Future directions include partial recoupling (bi-level or mixed spaces), dynamic re-planning in non-stationary environments (e.g., dynamic obstacle anticipation), and further exploitation of convex structure in hybrid energy-time optimal planning. Approaches that integrate rolling-horizon speed/path updates may partially address space–time conservatism without full complexity of joint space–time trajectory optimization (Zhou et al., 2020, Consolini et al., 28 Oct 2025, Li et al., 2021).

7. Summary Table: Exemplary Approaches and Key Attributes

Approach	Path Stage Technique	Speed Stage Technique	Real-Time Feasibility
FCP (Tariq et al., 6 May 2025)	Frenet nonlinear program	MPQP/Jerk-minimizing QP	0.03–0.07 s per plan
DL-IAPS+PJSO (Zhou et al., 2020)	SCP + polygonal bubbles	Piecewise-jerk QP	0.07 s end-to-end
Racing Planner (Kapania et al., 2019)	Iterative curvature minimization	Forward–backward speed integration	≲30 s for 4.5 km; ms–s (online preview)
Bezier+Trapezoidal (Li et al., 2021)	Static geometric path	Piecewise-Bezier QP (safe set)	6–10 ms/plan
Manipulator LP (Marauli et al., 23 Oct 2025)	Path precomputed offline	Parameter-linear LP or B-spline	>2× speedup vs. classical
Convex/DP Speed (Consolini et al., 28 Oct 2025, Consolini et al., 2021)	Path as input	Convex (LP/QP) / DP / SCA	O(n²); numerically efficient

Decoupled path-speed planning thus constitutes a technically mature, extensible framework, combining real-time feasibility with rigorous constraint handling across a diverse set of application domains.