Safety-aware Model Predictive Path Integral Control with Signal Temporal Logic

Technical Summary: Safety-aware Model Predictive Path Integral Control with Signal Temporal Logic

Problem Statement
Safety-aware motion planning in robotics faces significant challenges when missions are time-critical and governed by complex, real-time spatio-temporal specifications. While Signal Temporal Logic (STL) provides a rigorous formal language for expressing such requirements (e.g., "eventually reach a target while always avoiding obstacles"), existing planning methods struggle with computational efficiency. Traditional approaches using Mixed Integer Programs (MIP) offer rigorous guarantees but are often too expensive for online execution. Conversely, sampling-based methods like Model Predictive Path Integral (MPPI) are computationally efficient and parallelizable but typically lack formal guarantees for satisfying general STL constraints, often relying on cost-function approximations that do not directly map to safety satisfaction.

Methodology: Safety-aware-STL-MPPI
The paper proposes safety-aware-stl-mppi, a receding-horizon planning framework that integrates STL specifications directly into the MPPI controller using Control Barrier Functions (CBFs). The methodology proceeds in three main stages:

1. STL to Time-Varying CBF Encoding:
   The authors adapt the continuous-time CBF encoding for STL from prior work [5] to discrete-time systems. They construct a time-varying function $h_\phi(x, t)$ such that the condition $h_\phi(x, t) \geq 0$ implies the satisfaction of the STL formula $\phi$.

   • Recursive Construction: The encoding follows the recursive semantics of STL. Atomic predicates are mapped to smooth functions. Temporal operators (Globally $G$, Finally $F$, Until $U$) and logical conjunctions are handled by combining candidate CBFs using a smooth under-approximation of the minimum operator ($\tilde{\min}$).
   • Deletion Mechanism: To reduce conservatism, the framework employs a deletion strategy where CBF terms corresponding to temporal constraints that have already been satisfied or expired are removed from the optimization arguments at specific time steps.

2. MPPI with Safety-Augmented Costs:
   The core planner utilizes MPPI, which samples control input perturbations from a Gaussian distribution to generate candidate trajectories (rollouts).

   • Cost Function: The standard MPPI cost is augmented with a safety term derived from the CBF. The cost for a rollout $k$ is defined as $J^{(t)}_{\phi, k} = J^{(t)}_k + \eta \sum C_{CBF}$, where $C_{CBF}$ penalizes violations of the CBF condition (i.e., when the trajectory fails to maintain the forward invariance of the safe set).
   • Weighting: Samples are weighted based on their augmented costs to compute a nominal weighted rollout $v^+$.

3. CBF Projection (Safety Filter):
   Since the weighted rollout $v^+$ is not guaranteed to satisfy the CBF constraint, a projection step is applied. The algorithm solves a constrained optimization problem to find the closest control input $v^*_\tau$ to the nominal input $v^+_\tau$ that satisfies the discrete-time CBF condition:
   $$h_\phi(f(x_\tau, v^*_\tau), \tau+1) - h_\phi(x_\tau, \tau) \geq -\alpha h_\phi(x_\tau, \tau)$$

   • Approximation: Because the exact constraint is non-convex in discrete time, the authors propose a first-order Taylor expansion of the system dynamics and the CBF function. This transforms the constraint into an affine inequality, allowing the projection to be solved efficiently as a Quadratic Program (QP).
   • Closed-Form Solution: In the case of unbounded control inputs ($U = \mathbb{R}^m$), the QP admits a closed-form solution, further enhancing computational efficiency.

Key Contributions

• Framework Integration: The proposal of safety-aware-stl-mppi, a sampling-based planner that incorporates STL specifications as CBF-informed constraints while minimizing user-defined cost functions.
• Discrete-Time Adaptation: The adaptation of time-varying CBF encoding for STL from continuous-time literature to discrete-time systems, coupled with a lightweight projection mechanism to enforce these constraints.
• Empirical Validation: Demonstration of the algorithm's efficacy through numeric case studies and a quadcopter simulation in NVIDIA Isaac Lab.

Results
The authors evaluated the method against several MPPI baselines using four artificial Mars Rover planning case studies with diverse environments and cost setups.

• Safety and Efficiency: The proposed method consistently achieved high safety levels (satisfying the STL specifications) and maintained computational efficiency.
• Baseline Comparison: Other MPPI baselines, which typically treat STL satisfaction as part of the cost function without explicit constraint projection, failed to achieve the same level of safety guarantees.
• Hardware-in-the-Loop: A quadcopter planning experiment demonstrated the practical applicability of the approach.

Significance and Claims
The paper claims that safety-aware-stl-mppi bridges the gap between the computational efficiency of sampling-based planners and the rigorous safety guarantees of formal methods. The approach is grounded in an exact STL-CBF condition that provides formal sufficiency for satisfaction; however, the authors explicitly note that the nominal real-time implementation approximates this condition. Consequently, while the method aims to satisfy STL constraints, the formal guarantee (Theorem 2) applies strictly to controllers enforcing the exact constraint, whereas the proposed Algorithm 1 relies on a first-order approximation for real-time execution. The authors emphasize that their approach allows for general, non-convex cost functions while striving to ensure that the resulting closed-loop trajectory satisfies the user-defined STL mission requirements. The code for the algorithm is made publicly available to facilitate further research.