Experiments towards Safe Simulation-to-Reality Transfer of Learning-based Robot Controllers

Rapid prototyping of robotic controllers trained in domains different from their deployment environments presents numerous challenges, particularly regarding safety concerns. Typical laboratory test bench setups implement different forms of filters or control limitations to meet safety requirements. However, these are not always represented during the training phase, as their inclusion can adversely affect the learning process. To address these challenges, we developed a toolchain consisting of several ROS modules (Fig. 1). We evaluated its transferability from simulation to reality by testing different safety onfigurations during training. Our experiments explore how incorporating various safety constraints in simulation affects learning efficiency and the robustness of controllers when transferred to real-world applications. Our findings indicate that while including safety constraints during simulation does not necessarily slow down convergence or reduce peak performance, it significantly enhances the transferability of controllers to real-world scenarios where safety measures are mandatory. Conversely, omitting safety constraints during training and applying corrective mechanisms only at deployment can also yield effective outcomes.