Bridging the Simulation to Reality gap in robotics

Recent advances in Machine and Reinforcement Learning, particularly in visuomotor control policies for robotics, have increased reliance on simulation frameworks and physics engines. These tools generate synthetic data and create sandbox environments to meet the substantial data demands of neural network training. However, given the inherit discrepancies between simulation and reality, the Simulation to Reality (Sim2Real) Gap in Robotics refers to all factors and specialized techniques that affect a transfer of an agent from the simulation to the real-world. Our literature review revealed that this field is largely empirical, fragmented across the robotics landscape, and heavily influenced by technical aspects of visuomotor policy design. To address this, our methodology covers the Sim2Real domain comprehensively, establishing performance metrics, identifying Reinforcement Learning design considerations, and developing a taxonomy of specialized Sim2Real techniques. We also create a detailed taxonomy of available simulation frameworks and physics engines for robotics. The next phase of our research focuses on mushroom harvesting, an unsolved problem in industrial food automation. This interdisciplinary challenge involves complex kinodynamic task and motion planning under constraints and environment uncertainties related to deformable bodies and material failure modes. We develop a practical Sim2Real pipeline for mushroom harvesting using a robotic gripper, allowing us to evaluate several Sim2Real techniques, including system identification with modeling approximations and explicit transferable abstractions. Contrary to conventional Sim2Real approaches, we show that the simulation framework is not just a tool for training but should be an integral component of the perception and planning system. This is a key statement of the thesis, demonstrating the predictive power of simulation in real-world applications. Our concluding remark and future work directions, based on the experience gained during this work, point to a holistic point-of-view for active inference, where the robotic agent actions are point towards an active, life-long, real-world model discovery.