Synergistic Simplex: Cooperative Runtime Assurance for Safety-Critical Autonomous Systems

Technical Summary: Synergistic Simplex (SS)

Problem Statement
Autonomous systems, particularly autonomous vehicles (AVs), increasingly rely on machine learning (ML) for safety-critical tasks like perception and control. While ML offers high performance, it is inherently unverifiable and susceptible to long-tail faults (e.g., adversarial perturbations or missed detections) that can lead to catastrophic failures. Runtime Assurance (RTA) architectures, such as the Simplex and Perception Simplex (PS) designs, address this by pairing an unverified, high-performance ML layer with a verifiable, conservative safety layer. However, traditional RTA designs suffer from a major bottleneck: the safety layer operates in isolation, unable to utilize ML outputs. This leads to excessive conservativeness, where the safety layer overrides ML actions even when the ML output is safe but merely unverified, degrading system performance and efficiency. Furthermore, safety certification standards typically prohibit safety layers from using unverifiable ML information, creating a rigid separation that limits optimization.

Methodology
The paper proposes the Synergistic Simplex (SS) architecture, a generalization of the Simplex paradigm that enables bidirectional communication between the ML layer and the safety layer while preserving formal safety guarantees. The architecture consists of two primary communication links:

1. Safety-to-ML (S2M) Link: Instead of immediately overriding an unsafe ML action, the safety layer communicates verified constraints (e.g., obstacle locations) back to the ML layer. The ML layer is then requested to recompute an action within these verified bounds. If a safe recomputation is found, it is executed; otherwise, the safety layer applies a fallback safe action. This transforms the override mechanism from a hard switch into a cooperative process.
2. ML-to-Safety (M2S) Link: This link allows the safety layer to incorporate specific, selected outputs from the ML layer to refine decision-making. Crucially, the paper formalizes the conditions under which this is safe. Using a probabilistic fault propagation model on a Directed Acyclic Graph (DAG) of the autonomy stack, the authors prove that the safety layer can safely use ML outputs from independent functions (e.g., lane detection when monitoring obstacle detection) but must strictly prohibit the use of upstream functions (which could corrupt the monitored module) or downstream functions (which would create feedback loops violating the static fault model).

The authors provide a formal safety analysis proving that under specific assumptions (sound verification, safe maneuverable state space, and independence of used ML functions), SS preserves the deterministic safety guarantees of the original Perception Simplex.

Key Contributions

• Synergistic Simplex Architecture: A novel design enabling bidirectional interaction between ML and safety layers to improve performance without violating safety.
• Formal Safety Analysis: A rigorous derivation of conditions under which bidirectional communication preserves the safety guarantees of PS. Specifically, the paper proves that using independent ML outputs (like lane detection) in the safety layer does not compromise the safety guarantees regarding the monitored function (obstacle detection).
• Reference Design for AVs: A concrete instantiation of SS for autonomous vehicles, utilizing a LiDAR-based geometric safety detector and an ML-based perception stack (TransFuser++).
• Empirical Evaluation: Extensive evaluation in the CARLA and nuPlan simulation environments demonstrating the trade-offs between safety and performance.

Results
The evaluation focused on obstacle detection faults in the ML layer across two scenarios:

1. In-Lane Obstacle (Experiment 1): When an obstacle is missed in the ego lane, both PS and SS (M2S variant) successfully prevent collisions (Collision Rate = 0.0) by triggering emergency braking. However, in the CARLA setup, this results in a full stop and route failure (Route Completion Rate = 0.0). An auxiliary evaluation in nuPlan (supporting constraint injection) showed that the S2M link enables the planner to replan and avoid unnecessary stops, achieving a 74% route completion rate while maintaining safety.
2. Nearby-Lane Obstacle (Experiment 2): When an obstacle is missed in an adjacent lane, the standard PS architecture treats it as a critical threat, triggering conservative emergency braking and resulting in low route completion (10%). In contrast, the SS (M2S) variant uses ML-derived lane information to classify the obstacle as non-critical (Zone 2). This allows the system to apply lighter mitigation (throttle release) rather than emergency braking, significantly improving the route completion rate to 70% while maintaining zero collisions.

Significance and Claims
The paper claims that Synergistic Simplex successfully challenges the traditional strict separation between mission and safety layers. By formally deriving conditions for bidirectional communication, SS demonstrates that high-performance ML components and verifiable safety guarantees need not be competing objectives. The architecture allows safety-critical systems to leverage the performance benefits of ML (such as lane context) to reduce unnecessary conservativeness and improve task completion, all while retaining the deterministic safety guarantees of the underlying Simplex framework. The work posits that safety certification standards can be satisfied even when safety layers use ML information, provided the structural independence conditions are met.