Verification and validation of learning embodied agents

Develop verification and validation methodologies tailored to embodied agents that learn and adapt to novel experiences, providing meaningful safety assurances despite non-stationary, partially observable environments.

Background

Traditional verification and validation rely on fixed specifications and disturbance models that are difficult to obtain for adaptive, embodied agents operating in changing environments. Moreover, learning systems typically offer statistical guarantees that may not suffice for instantaneous safety claims.

The authors argue that new principles and approaches are needed to verify and validate learning embodied agents.

References

However, how to validate and verify the performance of an embodied agent that learns and adapts to novel experiences is an open question and it is likely that new principles for evaluating our embodied agents will be required.

— From Machine Learning to Robotics: Challenges and Opportunities for Embodied Intelligence  (2110.15245 - Roy et al., 2021) in Section 6.1 (Assessing Robot Learning: Verification and Validation)

The contribution is an admission-audit protocol with analytical and synthetic evidence; physical-robot and VLA validation remain open.

— When Validation Stops Learning: Auditing Update Admission for Continual Embodied Agents  (2609.10873 - Ma et al., 9 Sep 2026) in Abstract

The unresolved issue is validation. A world model can appear visually or metrically plausible while missing rare interactions, unusual agents, or causal effects of ego behavior.

— Planning-Oriented End-to-End Autonomous Driving: Architectures, Evaluation, and Emerging Paradigms  (2608.20111 - Guan et al., 20 Aug 2026) in Section 5.1, Why World Models Matter for Planning

Finally, while this framework defines the formal interface (the Cost Function) between agent health and swarm logic, the effectiveness of the reconfiguration depends on the convergence speed of the distributed planner. The verification and validation (V&V) of non-deterministic replanning algorithms remains an open research challenge.

— A Safety-Driven Architectural Framework for Fail-Operational Drone Swarms in Critical Missions  (2608.20906 - Giacomossi et al., 21 Aug 2026) in Section Discussion, subsection “Design Assumptions and Limitations”

When is there enough evidence to act? Safety-critical systems must avoid both underestimating long-tail risk and becoming paralyzed by unbounded hazard enumeration. The key question is whether available evidence supports a recoverable decision. Computation and caution should therefore adapt to consequence, epistemic uncertainty, recovery margin, and information value, while allowing the system to act, revise, sense, defer, or abstain.

— 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”