Seg-CURL: Segmented Contrastive Unsupervised Reinforcement Learning for Sim-to-Real in Visual Robotic Manipulation

Training image-based reinforcement learning (RL) agents are sample-inefficient, limiting their effectiveness in real-world manipulation tasks. Sim2Real, which involves training simulations and transferring to the real world, effectively reduces dependence on data. However, performance of transferred agent degrades due visual difference between two environments. This research presents a low-cost segmentation-driven unsupervised RL framework (Seg-CURL) solve Sim2Real problem. We transform input RGB views proposed semantic segmentation-based canonical domain. Our method incorporates levels Sim2Real: task-level transfers observation-level simulated U-nets segment scenes. Specifically, we first train contrastive RL(CURL) with segmented images simulation environment. Next, employ U-Nets robotic hand-view side-view during robot control. These U-Net pre-trained synthetic segmentation masks environment fine-tuned only 20 images. evaluate robustness both Seg-CURL is robust texture, lighting, shadow, camera position gap. Finally, our algorithm tested Baxter dark cube lifting task success rate 16/20 zero-shot transfer.