LyEvO framework enables safe robotics transfer from sim to real
Lyapunov-guided AI framework achieves 95% sim-to-real success rate in robotics tests
LyEvO, proposed by Riccardo Curcio, Hongpeng Cao, and Marco Caccamo, is a physics-grounded framework that combines constrained Evolutionary Optimization, Statistical Model Checking-based verification, and Lyapunov stability analysis. Using prior knowledge of system dynamics, it computes an initial candidate stability region, then iteratively optimizes and statistically verifies a policy while expanding that region based on verification results. Evaluated on Cartpole and 3D Quadrotor benchmarks through simulations and targeted real-world experiments, LyEvO demonstrates safe and robust sim-to-real transfer.
- LyEvO combines Lyapunov stability analysis with evolutionary optimization and statistical verification for safe sim-to-real transfer
- Achieved high success rates on Cartpole and 3D quadrotor benchmarks while maintaining safety guarantees
- Provides a systematic framework for deployment readiness assessment rather than trial-and-error real-world testing
Why It Matters
This framework could accelerate safe deployment of real-world robotics by replacing risky trial-and-error with mathematical guarantees