Learning with formal representations while preserving correctness guarantees

Identify learning methods that can be applied to formal representational languages in embodied systems while preserving guarantees of correctness needed for safety-critical inference and control.

Background

Formal logical and temporal specifications can offer correctness guarantees that are useful for safety in embodied systems. However, data-driven learning typically provides only statistical guarantees, potentially undermining formal safety assurances.

The authors seek approaches that combine learning with formal languages such that the ability to furnish correctness guarantees is retained.

References

There is an open question as to what kinds of learning can be applied to a formal representational language that preserves the ability to provide guarantees of correctness.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 4.2 (The Role of Logic: Opportunities and Future Directions)

How can learned consequences become enforceable constraints? RIWM may represent consequences in semantic or generative form, whereas runtime assurance typically operates on states, reachable sets, constraints, or verifiable temporal properties. Establishing a reliable interface between these representations while preserving uncertainty and applicability conditions remains a major challenge.

— Rethinking World Models for Safety-Critical Embodied Systems  (2609.03774 - Ma et al., 3 Sep 2026) in Section “Open challenges and outlook,” subsection “Open challenges”