Learning-based methods such as reinforcement learning (RL) and imitation learning (IL) offer new flexibility in robotics, enabling robots to handle complex tasks with reduced programming effort and mechanical constraints. However, the deployment of such systems in the real world introduces significant safety risks. To address these, we present Safe2Real – a modular toolchain designed to enable the safe and efficient transfer of trained control policies from simulation to physical robots.
The toolchain comprises three key components: (1) a simulation platform optimized for reality-gap reduction, (2) a ROS 2 software stack for policy deployment, perception, and logging, and (3) a unified safety controller available in both simulation and hardware environments. We evaluate Safe2Real using a cube-pushing task with a Franka Emika Panda robot and three training configurations with varying safety constraints. Our experiments show that integrating safety mechanisms during training leads to more robust and accurate real-world behavior without compromising learning performance. Particularly, including a safety controller during training significantly reduces sim-to-real mismatch, while even default policies can perform well when safety is enforced during execution. All implementations and resources will be made available via GitHub.