ScholarWorks@성균관대학교

초록

Sim-to-Real learning aims to deploy reinforcement learning (RL) policies trained in simulation directly to real robots. However, domain discrepancies between simulation and reality often lead to degraded performance. This study proposes a data-efficient framework that minimizes real-world trials while enabling risk-aware filtering of simulated policies' real-world performance. The proposed method integrates multi-metric policy scores obtained from simulation with a single batch of real grasp outcomes to build a Gaussian Mixture Model (GMM)-based surrogate that learns the simulation-reality mapping. The pipeline consists of three stages: (1) evaluating policies in domain-randomized simulators to extract feature and score vectors, (2) collecting a limited set of real-world calibration data, and (3) fitting the GMM surrogate to rank policies using log-likelihood and fused task scores. After brief calibration, the surrogate estimates the relative real-world reliability of untested policies without additional trials, prioritizing stability and worst-case avoidance over absolute peak performance under a constrained on-robot budget. Experiments on a controlled table-top grasping setup show improved Top-K deployment success and data-efficiency compared to simulation-only selection, while maintaining a fixed perception-to-candidate interface shared across simulation and real executions. Crucially, the framework functions primarily as a conservative safeguard that mitigates high-risk selections caused by unmodeled Sim-to-Real gaps, thereby providing a practical foundation for maintaining selection consistency in safety-critical or cost-sensitive deployments rather than pursuing iterative adaptation or retraining.

키워드