Learning-Based robot controllers often omit safety constraints during training, as their implementation can hinder the learning process and requires additional effort. We present a modular workflow to enable the safe deployment of control policies trained in simulation to physical hardware consisting of three key components: a standardized interface for exporting control models, a ROS 2 environment for deployment, and a safety controller that enforces joint-space limits consistently across simulation and real-world execution. We provide the first systematic evaluation of three practical safety formulations for highly parallelized training using canonical benchmarks to capture large- and small-action-space settings. Our experiments show that including a safety controller during training does not slow convergence or reduce peak performance, but significantly improves transferability for broad, free-space motion tasks. Although, training without safety constraints applying corrections only at deployment still proves effective for fine-grained assembly tasks. The implementation is available on GitHub.
Wrede et al. (Mon,) studied this question.