Efficient Receding-Horizon Solvers

• Efficient receding-horizon solvers are computational frameworks that decompose long-horizon optimal control, trajectory planning, and resource allocation tasks into short, moving-window subproblems with recursive feasibility.
• They employ layered architectures—combining fast nominal planning with nonlinear optimization—and leverage techniques like terminal cost bounds, the turnpike property, and hybrid relaxations to ensure stability and rapid convergence.
• Empirical outcomes demonstrate significant computational speed-ups and performance gains across various domains, including vehicle maneuvering, multi-agent monitoring, and humanoid robotics.

Efficient receding-horizon solvers are computational frameworks and algorithms designed to address optimal control, trajectory planning, state estimation, or resource allocation problems over a finite moving time window, with particular emphasis on reducing per-step computational burden for real-time implementation. Such solvers harness tailored mathematical structure, hybrid optimization formulations, model simplifications, event-driven execution, and parallelization to ensure recursive feasibility, stability, and significant speed-up relative to naively solving the full horizon problem at every update.

1. Architectural Principles of Efficient Receding-Horizon Solvers

Efficient receding-horizon approaches generally decouple long-horizon tasks into a series of small-scale, incrementally updated subproblems, using the most recent state or measurement as the initial condition and constraining the end-point via terminal costs or value-function estimates. A central innovation is the layered or two-step architecture, combining a fast but coarse combinatorial or sampling-based planner with a secondary local continuous optimizer over a receding window. For instance, the algorithm in (Bergman et al., 2019) employs an initial lattice-based nominal planner for wide-area feasibility, followed by a nonlinear receding-horizon refinement that incrementally optimizes a trajectory segment.

A schematic workflow for such two-step planners is:

1. Generate a nominal, feasible trajectory solving all combinatorial or long-horizon structural constraints.
2. At each receding-horizon iteration:
   • Fix the starting state and prescribe the endpoint along the previous best trajectory (the "terminal manifold").
   • Solve a nonlinear optimal control problem (often via direct multiple shooting or sequential quadratic programming) over a short window.
   • Augment the terminal cost with a cost-to-go bound from the previous nominal, ensuring recursive feasibility.
   • Stitch the solution with previously accepted segments and only accept the update if the global objective decreases.

Other architectures adopt event-driven local optimization (Welikala et al., 2020), sample-efficient tree search with subtree reuse (Lathrop et al., 2024), multi-fidelity relaxation (Wang et al., 2023), or adaptive horizon sizing coupled with stochastic search (Lukina et al., 2016).

2. Mathematical Formulations and Theoretical Guarantees

Efficient solvers leverage specific mathematical structures to enable speed without sacrificing performance:

• Terminal Manifold and Cost-to-Go Bounds: Efficient planners replace classical Lyapunov-terminal penalties with state-dependent upper cost-to-go bounds derived from the previous solution, as in (Bergman et al., 2019). This ensures that for each receding-horizon subproblem, an admissible upper bound on future cost is imposed, yielding recursive feasibility and convergence guarantees.
• Turnpike Property: In linear-quadratic and infinite-dimensional problems, efficient horizon reduction is achieved by exploiting the turnpike property—most of the optimal trajectory lies exponentially close to a static steady-state (the so-called "turnpike") (Breiten et al., 2018). With judicious terminal cost selection—typically the Lagrange multiplier of the static problem—exponential convergence to the infinite-horizon optimum is established, even with moderate prediction horizons.
• Branch-and-Bound and Relaxation for Hybrid/Integer Problems: For mixed-integer or hybrid systems, indirect approaches construct algebraic necessary conditions via the hybrid maximum principle, reducing each fixed discrete sequence to a system of algebraic equations (Tavassoli, 2018), while continuous relaxation of MINLPs into smooth nonlinear programs with exactness guarantees enables standard SQP solvers to handle discrete events efficiently (Li et al., 2021).
• Formal Guarantees: Most efficient solvers provide formal proofs of recursive feasibility, objective non-increase, and finite-time convergence by induction and proper stitching (e.g., Lemmas and Theorems 1–3 in (Bergman et al., 2019)). Adaptive stochastic-horizon methods use level-based Lyapunov arguments to prove almost-sure convergence in nonconvex landscapes (Lukina et al., 2016).

3. Computational Complexity and Real-Time Implementation

Efficient receding-horizon solvers are characterized by tailored complexity properties:

• Local Problem Size: By confining each optimization to a short moving window (of length $T$), the number of decision variables per step is linear in $T$, as opposed to the total horizon length (which is often orders of magnitude larger) (Bergman et al., 2019, Zhang et al., 2019).
• Warm-Starting and Window Shifting: Solutions are initialized using the shifted and truncated previous trajectory, reducing the number of optimization iterations needed per step (Bergman et al., 2019, Zheng et al., 2023).
• Event-Driven Activation: Controllers may re-solve only on occurrence of relevant events (e.g., agent arrival at a node), not at fixed intervals, which avoids unnecessary computations (Welikala et al., 2020).
• Complexity Scaling: Shown in Table 1:

Methodology	Per-Iteration Complexity	Empirical Runtime
Lattice+RHP (Bergman et al., 2019)	$O(N)$ (N = window discretization)	0.3–0.8 s per RHP iteration
Event-driven RHC (Welikala et al., 2020)	$O(|N_i|)$ per agent at event	μs-scale per event
Multi-fidelity RHP (Wang et al., 2023)	Varies, e.g., 0.37–0.9 s per cycle	95–99% cycles online
Drifting MPC (Foffano et al., 6 Apr 2026)	$O(M\cdot \mathrm{cost}_{\mathrm{fp}})$	30 ms per rollout (H=30)

• Load Reduction Techniques: Approaches such as warm-starting, soft-constraints, Gauss–Newton Hessian approximations and parallel linear algebra are pivotal for sub-second execution (Zheng et al., 2023).

4. Representative Algorithmic Strategies

4.1 Two-Step Motion Planning and Refinement

(Bergman et al., 2019) presents a two-layer architecture: a fast combinatorial planner (lattice-based with motion primitives and A* search) followed by receding-horizon nonlinear optimization. Empirically, this reduces truck-and-trailer parking trajectory cost by 23–41% with sub-second updates. Acceptance is strictly tied to global objective non-increase criteria, utilizing cost-to-go upper bounds constructed from the previous solution as a terminal cost.

4.2 Event-Triggered and Distributed Control

(Welikala et al., 2020) introduces an event-driven, distributed persistent-monitoring RHC, where each agent solves small, closed-form rational fractional programs for local scheduling/coverage decisions upon event observation, yielding globally optimal actions at O(1) computational cost per regime. This structure underpins nearly 50% reduction in monitoring cost compared to parametric and centralized schemes.

4.3 Sample-Efficient Tree-Search Reuse

Model Predictive Trees (MPT) (Lathrop et al., 2024) hot-start the search process with an entire retained subtree, exploiting contraction-theoretic tracking to maximize search reuse across timesteps under model drift. Theoretical analysis bounds tracking error (as a function of tree depth, model error, and contraction rate), and empirical results show an order-of-magnitude sample efficiency gain versus cross-entropy and non-reuse tree-search approaches.

4.4 Multi-Fidelity and Locally-Guided Planning

(Wang et al., 2023) employs multi-level model fidelity, splitting the receding horizon into short, accurate execution windows and longer convex-relaxed prediction regions to rapidly estimate value functions with controlled suboptimality. A learned, locally-guided terminal cost further collapses the problem to near one-step optimization, achieving 95–99% real-time cycle-wise online execution in humanoid locomotion scenarios.

4.5 Adaptive-Horizon and Particle-Swarm Hybridization

ARES (Lukina et al., 2016) fuses receding-horizon synthesis with particle-swarm optimization, adaptively resizing both swarm and horizon using importance-splitting inspired level intervals. This guarantees convergence in nonconvex problems (e.g., V-formation flocking) at a fraction of the computation per level compared to fixed-horizon, fixed-size approaches.

5. Application Domains and Empirical Outcomes

Efficient receding-horizon solvers have demonstrated impact across domains, as summarized below:

• Large-scale Vehicle Maneuvering: Truck-and-trailer parking via two-step lattice+RHP planning (Bergman et al., 2019): 20–41% improvement in trajectory objectives, subsecond iteration times.
• Distributed Multi-Agent Monitoring: Event-driven RHC (Welikala et al., 2020): 40–60% monitoring cost reduction in multi-agent network settings.
• Nonprehensile Manipulation and Hybrid Dynamics: MPT (Lathrop et al., 2024) delivers 29.9% higher cumulative reward than cross-entropy with hotstart, enabling non-trivial recontact behaviors.
• Humanoid Locomotion: Locally-guided short-horizon RHP (Wang et al., 2023) achieves >95% online planning success with <0.4 s per update in both simulation and real robotic deployment.
• Electric Vehicle Speed and Gear Optimization: Continuous relaxation for MINLP gearshift (Li et al., 2021) yields 7–19% energy savings with solve times below 1–2 s for typical horizons.

6. Implementation Trade-offs and Practical Considerations

Efficient receding-horizon algorithms feature characteristic trade-offs:

• Window Length: Longer horizons yield diminishing gains after certain thresholds (3% or less past T≈60–80 s in (Bergman et al., 2019)) while increasing computational overhead.
• Terminal Cost Choice: Accurate terminal costs or cost-to-go bounds dramatically impact convergence rate and feasibility domains.
• Parallelization and Warm-Start: Key for handling high-dimensional systems at real-time rates.
• Relaxation vs. Integer Handling: Smooth relaxation strategies (e.g., simplex-based embedding for gear-shift) not only permit NLP solvers but also retain the optimal solution structure (Li et al., 2021).
• Subtree Reuse and Resetting: For sample-based planners, subtree retention is bounded by tracking error dictated by model mismatch and contraction rate. Hyperparameters (horizon, rollout count, reset threshold) should be chosen to balance sampling efficiency against tracking safety (Lathrop et al., 2024).
• Domain-Specific Constraints: Incorporation of spatiotemporal soft-barriers in dense traffic ensures proactive collision avoidance at negligible extra computational cost (Zheng et al., 2023).

7. Summary Table: Key Methods and Quantitative Performance

Solver/Reference	Core Strategy	Empirical Performance/Results
(Bergman et al., 2019)	Lattice+RHP two-step + cost-to-go bound	−23–41% cost, subsecond iteration
(Lathrop et al., 2024)	UCT tree search with subtree reuse	Order-of-magnitude sample speedup
(Welikala et al., 2020)	Event-driven closed-form RHC	50% objective reduction, μs runtime
(Wang et al., 2023)	Multi-fidelity/learned value function	95–99% cycle-wise online, <0.4 s/cycle
(Li et al., 2021)	Relaxed MINLP for BEV speed/gear	7–19% energy saving, 0.7–1.8 s/step
(Foffano et al., 6 Apr 2026)	Drifting MPC: KL-regularized generator	≈30 ms/rollout, near-oracle cost

• (Bergman et al., 2019) An Optimization-Based Receding Horizon Trajectory Planning Algorithm.
• (Breiten et al., 2018) On the Turnpike Property and the Receding-Horizon Method for Linear-Quadratic Optimal Control Problems.
• (Welikala et al., 2020) Event-Driven Receding Horizon Control For Distributed Persistent Monitoring in Network Systems.
• (Lathrop et al., 2024) Model Predictive Trees: Sample-Efficient Receding Horizon Planning with Reusable Tree Search.
• (Wang et al., 2023) Online Multi-Contact Receding Horizon Planning via Value Function Approximation.
• (Li et al., 2021) Coordinated Receding-Horizon Control of Battery Electric Vehicle Speed and Gearshift Using Relaxed Mixed Integer Nonlinear Programming.
• (Zheng et al., 2023) Spatiotemporal Receding Horizon Control with Proactive Interaction Towards Autonomous Driving in Dense Traffic.
• (Lukina et al., 2016) ARES: Adaptive Receding-Horizon Synthesis of Optimal Plans.
• (Tavassoli, 2018) An indirect computational procedure for receding horizon hybrid optimal control.
• (Zhang et al., 2019) Receding Horizon Optimization for Energy-Efficient UAV Communication.
• (Foffano et al., 6 Apr 2026) Receding-Horizon Control via Drifting Models.