Sim-to-Real Neural Learning with Domain Randomisation for Humanoid Robot Grasping

Collecting large amounts of training data with a real robot to learn visuomotor abilities is time-consuming and limited by expensive robotic hardware. Simulators provide a safe, distributable way to collect data, but due to discrepancies between simulation and reality, learned strategies often do not transfer to the real world. This paper examines whether domain randomisation can increase the real-world performance of a model trained entirely in simulation without additional fine-tuning. We replicate a reach-to-grasp experiment with the NICO humanoid robot in simulation and develop a method to autonomously create training data for a supervised learning approach with an end-to-end convolutional neural architecture. We compare model performance and real-world transferability for different amounts of data and randomisation conditions. Our results show that domain randomisation improves the transferability of a model and can mitigate negative effects of overfitting.