- The paper introduces a smooth, numerically exact parameterization of CT-STL for trajectory optimization, eliminating conservative, discrete approximations.
- It employs multiple-shooting discretization with first-order hold, enabling dense-time evaluation of STL predicates and improved gradient flow.
- Empirical results demonstrate millisecond-level solution times and scalability, with strong potential for UAVs, autonomous vehicles, and robotic applications.
Smooth and Exact Dense-Time STL Parameterization for Optimal Control
Introduction
This work (2604.04245) addresses the persistent challenges in integrating complex continuous-time Signal Temporal Logic (CT-STL) specifications into nonconvex trajectory optimization frameworks. Existing approaches predominantly leverage discrete-time approximations or introduce auxiliary constructs (e.g., control barrier functions, specialized planners) that lead to suboptimal conservatism, limited temporal expressiveness, or prohibitively expensive mixed-integer encodings. The authors propose a smooth, numerically exact parameterization of CT-STL evaluated in dense time, facilitating direct embedding into gradient-based optimal control pipelines. This approach generalizes recent discrete-time generalized mean-based smooth robustness (D-GMSR) semantics to the continuous domain, providing an operational framework for handling rich temporal-logical tasks โ including until and eventually operators โ while guaranteeing continuous-time satisfaction and computational tractability.
CT-STL Semantics and Optimization Framework
Signal Temporal Logic predicates are defined over real-valued signal trajectories, enabling the direct encoding of time-varying and logical requirements such as path, safety, and reachability constraints. The classical quantitative semantics, built from non-differentiable min/max operations, hinder convergence and robustness in gradient-based solvers and suffer from locality/masking. The authors transplant GMSR semantics onto dense-time representations, enabled by a multiple-shooting discretization and first-order hold (FOH) input parameterization, ensuring that the differentiable robustness measure depends smoothly on the entire trajectory.
Key features of the proposed framework include:
- Direct dense-time evaluation of always-type predicates: Ensures no time intervals are undersampled, critically important for path and obstacle avoidance constraints.
- Removal of node-induced conservatism in eventually-type predicates: Satisfaction may occur at arbitrary times within the allowed interval, as opposed to fixed nodes.
- Support for general STL formulas: Until and composite specifications are handled recursively using smooth logical and temporal compositions at the discretized subnode level.
- Well-scaled, non-masked, and non-local gradient information: The robustness landscape constructed by GMSR semantics has improved properties for gradient-based optimization, as gradients are distributed among multiple time points and predicates, ameliorating the discontinuity and masking of min/max-based forms.
The trajectory optimization problem is thus formulated as a finite-dimensional nonlinear program, where system dynamics are enforced at discretization nodes, and STL robustness and feasibility is evaluated on a dense subgrid aligned with the numerical integration scheme.
Numerical Implementation and Empirical Results
A representative feasibility problem is presented involving a quadrotor platform. The task incorporates both path constraints (tilt, thrust, velocity, encoded as always-type G operators) and a non-trivial until-type constraint, requiring low speed enforcement until a specified charging station is reached. The smooth robustness semantics are evaluated on RK4-integrated multiple shooting states, and enforced via exact-penalty minimization in a prox-convex framework. The authors report:
- Dense-time path constraint enforcement: All constraints are continuously satisfied on the fine integration grid.
- Correct logical behavior: The vehicle remains below the specified speed until entering the charging station, with the time of entry exactly identified by the robustness witness.
- Computational efficiency: Example problems are solved in milliseconds (e.g., ~15ms for all convex subproblems on a commodity laptop with non-optimized code), and the approach permits much coarser optimization grids due to dense-time evaluation, reducing the overall program dimensionality.
- Practical feasibility/correctness: Final solutions show no violations of path or STL specifications and are dynamically consistent to tolerance.
Theoretical Properties and Numerical Advantages
The key theoretical contribution is the sound and complete parameterization of CT-STL on the numerically integrated trajectory (to within the integration accuracy). Unlike prior MILP or CBF-based approaches, the method handles nonlinear systems and generic STL formulas without the need for user-side construction of invariance, tracking bounds, or mixed-integer encodings.
The generalized mean-based parameterization regularizes near the decision boundaries, ensuring C1 smoothness and eliminating singular Jacobian behavior. Logical and temporal operators in the dense-time domain further ensure that the desired robustness margin propagates throughout all relevant subformulas and over all relevant time samplesโa substantial improvement over the min/max-centric approaches, which typically provide gradient information only at a single worst-case witness time or predicate (leading to slow or stalled convergence in flat robustness regions).
Implications and Future Directions
Practical implications are significant for high-assurance and safety-critical application domains where continuous-time STL constraint satisfaction is non-negotiable (UAV guidance, autonomous vehicles, and robotic motion planning). The framework's ability to scale to coarser temporal grids without loss of logical fidelity directly translates to reduced optimization problem size and improved computational tractability. The STL-in-the-loop approach, as implemented, is agnostic to system dynamics and STL formula structure, with direct applicability to emerging differentiable planning and learning systems.
Theoretical implications include the removal of key obstacles to tightly coupling dense-time STL constraint satisfaction to modern nonconvex optimal control solvers. This opens the path toward more general integration of logical-temporal requirements in direct collocation and convexification approaches, and forms a basis for further extensions: e.g., free-final-time control with STL constraints or dynamic augmentation so that STL robustness is treated as an explicit state variable.
Conclusion
This work establishes a robust, semantically exact, and computationally efficient framework for embedding arbitrary CT-STL specifications into nonconvex trajectory optimization, enabled by a smooth and dense-time GMSR parameterization. The method unifies logical and path constraint handling, eliminates locality and masking, and is readily implemented using standard multiple-shooting and modern automatic differentiation toolchains. Subsequent theoretical and practical advancements are expected from extending the framework to free-final-time problems and more general dynamic augmentation, further strengthening the integration of temporal logic with optimal control and AI-based planning architectures.