LyEvO: A New Framework for Safe and Robust Sim-to-Real Policy Learning

A new framework called LyEvO has been developed by researchers to train controllers that ensure safety and reliability in simulations, while also evaluating their suitability for real-world applications. This framework merges constrained Evolutionary Optimization with Statistical Model Checking (SMC) for verification, alongside Lyapunov-based stability assessments. By utilizing existing knowledge of system dynamics, it first identifies a candidate stability region, then iteratively refines this region by optimizing and statistically verifying a policy using operational scenarios. This comprehensive method establishes a practical standard for determining deployment readiness. The framework underwent rigorous testing on Cartpole and 3D Quadrotor benchmarks through simulations and specific real-world trials. It addresses significant challenges in sim-to-real transfer, providing a structured method for confirming safety and robustness prior to deployment. The research can be found on arXiv with the identifier 2608.06481.