Realeasy: real-time capable simulation to reality domain adaptation

We address the problem of insufficient quality of robot simulators to produce precise sensor readings for joint positions, velocities and torques. Realistic simulations of sensor readings are particularly important for real time robot control laws and for data intensive Reinforcement Learning of robot movements in simulation. We systematically construct two architectures based on Long Short-Term Memory to model the difference between simulated and real sensor readings for online and offline application. Our solution is easy to integrate into existing Robot Operating System frameworks and its formulation is neither robot nor task specific. The collected data set, the plug-and-play Realeasy model for the Panda robot and a reproducible real-time docker setup are shared alongside the code. We demonstrate robust behavior and transferability of the learned model between individual Franka Emika Panda robots. Our experiments show a reduction in torque mean squared error of at least one order of magnitude.