Transferring Safety Guarantees from Abstract Models to Concrete Systems

16 concepts9 papers

Formally proving that a physical system—such as an autonomous vehicle, a medical device, or a walking robot—is safe is an incredibly difficult challenge. Physical hardware operates in a high-dimensional, continuous world filled with non-linear dynamics, measurement noise, and unmodeled disturbances. Performing formal verification directly on these complex systems is often computationally intractable.

To bypass this complexity, engineers typically verify safety properties or design safe controllers on simplified, low-dimensional abstract models. However, a critical question remains: Why should a safety proof valid for an idealized model hold true for a messy physical system?

To bridge this gap, a powerful paradigm has emerged around safety property transfer. By establishing rigorous mathematical relationships between abstractions and concrete implementations, formal safety guarantees can be reliably propagated from the model to the real world.

Reachset Conformance: The Envelope Principle

The most direct way to transfer safety from a model to a physical system is to ensure that the model's mathematical predictions act as a conservative "envelope" over reality. This concept is formalized as reachset conformance.

As established by Roehm et al., reachset conformance is a necessary and sufficient condition for transferring safety properties. The core intuition is straightforward: if the set of all states reachable by the abstract model (the reachable set) completely encloses all possible trajectories of the real system, then any unsafe state avoided by the model must also be avoided by the physical system.

  [ Unsafe State Space ]
  ----------------------
     _________________
    /  Abstract       \
   /   Reachable Set   \
  /     ___________     \
 /     /  Real     \     \
 |     | Trajectory|     |
 \     \___________/     /
  \                     /
   \___________________/

Rather than requiring a perfect, mathematically precise match between the model and reality—which is virtually impossible to achieve—this framework embraces uncertainty. Liu et al. demonstrate that reachset conformance relaxes strict classical simulation relations. This makes it highly practical for real-world robotics, as engineers can identify models from experimental data and guarantee safety as long as the model's predicted reachable sets encompass all observed behaviors.

To implement this in practice, Althoff introduces algorithmic tools in Checking and Establishing Reachset Conformance in CORA 2023 that automatically check experimental data against a model's predictions. If the real system drifts outside the envelope, the tool synthesizes the minimum necessary non-determinism (uncertainty bounds) to expand the reachable set until it fully contains the physical behavior, restoring the validity of the safety transfer.

Bounding the Mismatch: Simulation and Barrier Functions

When designing continuous controllers, systems are often modeled using a low-dimensional Reduced-Order Model (ROM) to plan trajectories, which are then executed by a high-dimensional Full-Order Model (FOM). Because the full system has dynamics that the simplified model ignores, the real system will inevitably drift from the abstract plan.

To transfer safety in these scenarios, researchers use simulation functions to act as a mathematical tether between the two levels of complexity.

Cohen et al. formalize this by constructing a Control Barrier Function (CBF) for the full-order system that explicitly combines the ROM's safety conditions with a simulation function. The simulation function quantifies and bounds the maximum tracking error between the ROM and the FOM. By padding the abstract safety boundaries with this error bound, the full-order system is guaranteed to remain within a safe "tube" around the abstract trajectory.

Similarly, Lutkus et al. introduce the concept of approximate dynamics conjugacy to map safety and stability certificates from low-dimensional latent spaces back to high-dimensional systems. By rigorously bounding the dynamical mismatch between the latent space and the true space, Lyapunov and barrier functions designed in the simplified space retain their formal guarantees when deployed on the full system.

This tight coupling ensures what Kang et al. call persistent feasibility: if a safe control action can be found within the simplified abstract model, a mathematically valid, safe counterpart is guaranteed to exist for the real-world concrete system.

Hierarchical Refinement and Data-Driven Abstractions

Safety property transfer is equally vital when systems are composed dynamically or when the underlying physics are completely unknown.

In medical cyber-physical systems (MCPS), such as a network of ventilators and infusion pumps assembled at a patient's bedside, safety must be verified on the fly. King et al. propose a framework where safety properties verified on an abstract model are transferred to the physical setup via a trusted software platform. The platform acts as a safety enforcer, verifying that the physical device configurations strictly refine the constraints of the abstract model before deployment. This hierarchical verification strategy is echoed by Murugesan et al., who bridge different modeling formalisms to ensure that detailed, concrete component designs preserve the safety properties proven at higher levels of system abstraction.

But what if we do not have a reliable first-principles model to abstract? The authors of Formal Verification of Unknown Dynamical Systems via Gaussian Process Regression address this by combining data-driven learning with formal methods. They use Gaussian Processes to learn unknown system dynamics from input-output data alongside rigorous statistical confidence bounds. These bounds are used to construct an Interval Markov Decision Process (IMDP) abstraction. Because the abstraction accounts for the worst-case learning error, any verification results obtained on the discrete IMDP abstraction can be formally mapped back to provide ironclad safety guarantees for the original, unknown dynamical system.

Summary

Safety property transfer works because it replaces the impossible task of verifying a complex physical system with a two-step process:

  1. Verify the Abstraction: Prove safety within a simplified, computationally compliant mathematical sandbox.
  2. Quantify the Gap: Use reachset conformance, simulation functions, or refinement mappings to guarantee that the real system cannot deviate from the abstraction in a way that violates safety.

By separating the verification of safety logic from the management of physical complexity, this paradigm provides a scalable blueprint for building provably safe autonomous and cyber-physical systems.

Key Foundational Work

Assuring the safety of on-demand medical cyber-physical systems (2013) — contributed 3 concept elements to this direction.
Linking abstract analysis to concrete design: A hierarchical approach to verify medical CPS safety (2014) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_safety_property_transfer_3dff9fcd,
  title = {safety property transfer},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/safety-property-transfer/rdr_3dff9fcd358149b59c45b11b3e8aafd5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_safety_property_transfer_3dff9fcd_v1,
  title = {safety property transfer},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/safety-property-transfer/rdr_3dff9fcd358149b59c45b11b3e8aafd5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}

See something wrong? Report a mistake