Sim-to-Real reinforcement learning aims to deploy policies trained in simulation to real robots fast and safely; however, discrepancies in physics, sensing, and contact conditions introduce a domain gap that often degrades real-world performance. To m...
Sim-to-Real reinforcement learning aims to deploy policies trained in simulation to real robots fast and safely; however, discrepancies in physics, sensing, and contact conditions introduce a domain gap that often degrades real-world performance. To mitigate this issue, this thesis proposes \textit{Illy-Net}, a data-efficient surrogate-based policy ranking framework. Illy-Net (i) collects performance logs through domain-randomized evaluations in Isaac Sim, (ii) learns a GMM surrogate from a \emph{single batch} of real-world trials to calibrate and predict real-world compatibility, and (iii) optionally integrates PPO (discrete candidate selection) and SAC (continuous pose refinement) to compute the final deployment ranking. The final score is computed as $X_i=s_i^{\alpha}r_i^{1-\alpha}$ or $X_i=q_i^{\beta}r_i^{1-\beta}$, and only the Top-$K$ candidates are deployed and validated on a UR16e+RG6 platform.
The experimental pipeline generates and normalizes grasp candidates using SAM-6D and GraspNet, evaluates candidates under randomized factors (e.g., illumination, friction, mass, pose/sensor noise), and estimates real-world compatibility using a GMM trained from a single batch of real data. For short-range transfer in a logistics packaging task, a no-drop criterion is adopted as the primary success condition. Across seven objects, Illy-Net reduces MAPE from 38.2\% to 8.1\% and improves the Top-5 success rate from 77.3\% to 91.7\%, while reducing the required real-world trials from 250 to 65 (a 74\% reduction). We are developing a candidate generation module that searches for parallel plane pairs via surface segmentation and synthesizes grasp candidates by aligning surface normals with the gripper geometry; the generated candidates will be used for simulation training. Overall, combining large-scale simulation exploration with single-batch calibration narrows the Sim-to-Real gap, reducing experimentation cost/time and deployment risk in industrial settings.