learnINg versaTile lEgged locomotioN wiTh actIve perceptiON

移动机器人技术有望在未来几年得到快速发展。设计高效的移动机器人平台将在不同的场景中打开各种应用程序，并可能在繁重、重复或危险的任务中取代人类。在所有移动机器人平台中，腿部机器人尤其令人感兴趣。它们的结构使它们能够以不同于需要平坦地面的轮式平台的方式导航具有挑战性的区域，例如，当地面不规则、存在碎片或在狭窄的空间中，并携带沉重的有效载荷，与其他生物启发的方法不同。该项目的目标是开发设计高度自主的腿式平台的基本方法。所需的解决方案应适应低间隙和崎岖地形的苛刻环境，使机器人能够避免主动故障状态并自主恢复。为了显著提高步态控制器的自主性，我们需要更深入地研究环境相互作用的一些基本方面：感知和动作。我们将利用现代机器学习技术来创建更通用的框架的理论和实践基础。我们方法的主要新奇之处在于对感知和行动的学习进行联合分析。事实上，有许多重要的理由需要共同考虑行动和感知。研究人员证明，主动和交互感知是提高感知和估计无法通过被动感知环境测量的数量的能力的基础，特别是在操纵任务中。此外，为了在感知和运动学习中建立强大的算法基础，关键是要考虑每个模块产生什么类型的输入和输出数据，如何处理数据，以及哪种表示更适合学习过程。这一集成的解决方案将使我们的系统能够感知周围环境，对意外的地形特性做出反应，避免滑动和故障，感知与周围障碍物的接触，并利用这种物理交互来高效地通过狭窄的通道，超越了当前最先进的方法。我们研究的最终目标是利用高质量的科学评估标准和基准来验证方法论。为了更好地突出开发的解决方案的优点和最终的缺点，我们将在洞穴探索场景中测试建议的系统。我们的目标是展示该解决方案在超越现有移动测试场景的任务中的有效性，并提供一种简单的方法来将我们的平台与不同的商业硬件和软件解决方案进行比较。

Mobile robotics is expected to have a rapid development in the following years. The design of efficient mobile robotics platforms will open various applications in different scenarios and could substitute humans in laborious, repetitive, or dangerous tasks. Among all mobile robotics platforms, legged robots are of particular interest. Their structure allows them to navigate challenging areas differently from wheeled platforms that require flat ground, e.g., when the ground is irregular, in the presence of debris, or in tight spaces, and carry heavy payloads, differently from other bioinspired approaches. The objective of this project is to develop the fundamental methodology to design highly autonomous legged platforms. The desired solution should adapt to demanding environments with low clearance and rough terrain, enabling the robot to avoid actively faulty states and recover autonomously. To significantly improve the autonomy of the gait controllers, we need to have a deeper look at some fundamental aspects of the interaction of the environment: perception and action. We will exploit modern machine learning techniques to create a more general framework's theoretical and practical foundations. The key novelty of our approach lies in the joint analysis of learning for perception and action. Indeed, there are many important reasons to consider action and perception jointly. Researchers demonstrated that active and interactive perception is fundamental to improving the ability to perceive and estimate quantities that cannot be measured by passively sensing the environment, particularly in manipulation tasks. Furthermore, to build strong algorithmic foundations in perception and locomotion learning, it is crucial to consider what type of input and output data is produced by each module, how the data is processed, and which kind of representation is more suitable for the learning process. This integrated solution will enable our system to be aware of the surrounding environment, react to unexpected terrain properties avoiding slippage and failures, sense contact with the surrounding obstacles, and exploit this physical interaction to locomote efficiently through narrow passages, going beyond current state-of-the-art methods. The final objective of our research is to verify, employing high-quality scientific evaluation standards and benchmarks, the methodology. To better highlight the developed solution's strengths and eventual drawbacks, we will test the proposed system in a cave exploration scenario. Our objective is to demonstrate the solution's effectiveness on a task that goes beyond existing locomotion test scenarios and provide an easy way to compare our platform with different commercially available hardware and software solutions.