Certifiable Safe RL Framework Reduces Robot Conservatism

A new paper accepted to IROS 2026 presents a certifiable safe reinforcement learning framework that addresses a key limitation in existing safe RL methods: overly conservative robot behavior due to neglecting control-affine structure in residual model uncertainties. The authors—Hao Zhou, Yanze Zhang, Cameron Reid, and Wenhao Luo—introduce an approach that first uses Control-Affine Random Fourier Features (ARFF) to learn robot dynamics in a control-affine form. This offers computational efficiency that scales with dataset size and reduces model bias. Then, adaptive conformal prediction (ACP) provides model-free, efficient uncertainty quantification for the safety constraints derived from the learned dynamics.

This uncertainty quantification is integrated with control barrier functions (CBF) to synthesize a safe, data-driven policy with certifiable guarantees. The framework avoids the need for prior knowledge of nominal dynamics or overly conservative safety margins. Simulation results on cartpole and 3D quadrotor platforms demonstrate that the method achieves task efficiency while maintaining provable safety. By structuring the learned dynamics and quantifying uncertainty rigorously, the work bridges the gap between theoretical safety guarantees and practical robot learning.