Antonio Loquercio – författare

Transfer simulation-trained AI to real-world robotic platforms effectively Training AI in simulation offers efficiency and safety advantages, but deploying that intelligence on physical platforms introduces challenges that can undermine performance. Artificial Intelligence: From Simulation to Reality addresses this critical gap directly. Compiled by researchers from DARPA, Johns Hopkins, Penn, and Oregon State, this volume provides the methodologies needed to successfully transfer simulated learning to real-world autonomous systems. The book covers diverse simulation environments , AI techniques for sim-to-real transfer, and a variety of exciting and relevant application domains, including autonomous vehicle drifting, bipedal locomotion, control of humanoid robots, human-in-the loop robotics, quadruped autonomy, and superhuman drone racing. This book also presents a modern treatment of classical concepts in robotics, including how large language models and vision-language-action model training techniques can be adapted to train robots in simulation for real-world transfer. Each chapter addresses specific sim-to-real challenges with proven solutions. Readers will also explore: Coverage of multiple simulation platforms, environments, and techniques enabling practitioners to select the right tools for their specific robotics applicationsDomain randomization and system dynamics techniques that improve the robustness of AI models when transitioning from simulated to physical environmentsMachine learning methods for predicting robot trajectories that account for real-world uncertainties absent from idealized simulation training scenariosTechniques adapted from large language model development showing how transformer-based approaches can enhance sim-to-real transfer for autonomous robotsPractical guidance on addressing the quintessential challenges that arise when deploying simulation-trained intelligence on real autonomous platformsResearch scientists, applied scientists, and engineers working in AI, machine learning, or robotics will find this an authoritative resource for sim-to-real transfer. Professors teaching robotics, transfer learning, reinforcement learning, or AI for control courses will find material suitable for advanced undergraduate and graduate curricula.

This book presents the astonishing potential of deep sensorimotor policies for agile vision-based quadrotor flight. Quadrotors are among the most agile and dynamic machines ever created. However, developing fully autonomous quadrotors that can approach or even outperform the agility of birds or human drone pilots with only onboard sensing and computing is challenging and still unsolved.Deep sensorimotor policies, generally trained in simulation, enable autonomous quadrotors to fly faster and more agile than what was possible before. While humans and birds still have the advantage over drones, the author shows the current research gaps and discusses possible future solutions.