RoboHike: Autonomous Quadrupedal Robot Navigation and Hiking in Challenging Rough Terrains

四足机器人正在获得重要的能力，特别是在过去的十年中，由于机电一体化，控制和规划的快速发展。在机器人需要检查难以到达的区域或在危险环境中帮助人类的情况下，四足机器人可能是理想的，因为它们能够以安全和节能的方式处理稀疏的立足点。迄今为止，四足机器人能够通过一些类型的崎岖地形，通常使用传统的控制和感知方法。然而，它们的移动性仍然远远落后于它们的自然对应物，特别是在环境动态变化的情况下。诸如导航和在崎岖或多岩石的小径上行走等任务，其环境本身是不确定的，不能完全感知，并且可能动态变化，这仍然是腿式机器人运动的核心挑战。RoboHike旨在为四足机器人引入和开发新颖的高层次和平台无关的感知和学习方法，用于建模、识别和绘制立足点，从而有可能在具有挑战性的地形上实现快速、稳健和可靠的导航和徒步技能。特别是，它旨在结合各种传感系统，如本体感觉（例如，惯性，速度或关节扭矩）和外感觉（例如，视觉，范围，事件或脚的力接触数据）感知来重建环境和处理潜在丢失或不准确数据的不确定性，在运动之前和运动期间，特别是动态变化的地形。这将使新的脚步规划和机器人在环境中的定位成为可能。分析和（自我监督和强化）学习方法将利用多模态传感，让四足机器人模仿动物在学习走路时规划脚步的方式。开发的方法将在几个全尺寸四足机器人上进行实验验证，用于学术和工业实际用例，用于检查，巡逻和维护等任务。RoboHike将致力于在建筑领域、石油和天然气领域、或人为/自然灾害后受损的领域开发下一代自主机器人系统，这些领域需要高效导航，并且可能存在崎岖/岩石地形、工业楼梯、管道和狭窄通道。其愿景是赋予四足动物环境认知能力，使其能够自主从事艰苦或危险的体力劳动。预计对国家和工业部门的自动检查、监测、维护和灾害管理的影响将很大，这些部门的地形复杂，及时干预的要求至关重要。我们打算在具有挑战性的赛道上构建公开共享的基准数据集，通过使机器人能够在具有挑战性的场地上工作，从而使机器人社区在机器人运动方面向前迈进了几步。

Quadrupedal robots are gaining important capabilities, especially over the past decade, due to the rapid advancements in mechatronics, control, and planning. In scenarios that robots need to operate for either inspecting hard-to-reach areas or aiding humans in dangerous and hazardous environments, quadrupedal robots could be ideal due to their ability to deal with sparse footholds in a safe and energy efficient way. To date, quadrupedal robots are able to traverse some types of rough terrain, using usually traditional control and perception methods. However, their mobility is still far behind their natural counterparts, especially in cases that the environment is dynamically changing. Tasks such as navigating and hiking rough or rocky trails, where the environment itself is uncertain, not fully perceived, and potentially dynamically changing, remain central challenges in legged robot locomotion. RoboHike aims at introducing and developing novel high level and platform-agnostic perception and learning approaches for modeling, identifying, and mapping footholds for quadrupedal robots, such that it would be possible to achieve fast, robust, and reliable navigation and hiking skills on challenging terrains. In particular, it aims at combining various sensing systems, such as proprioceptive (e.g., inertia, speed, or joint torques) and exteroceptive (e.g., visual, range, event, or foot's force contact data) perception to reconstruct the environment and handle the uncertainty of potentially missing or inaccurate data, before and during locomotion, especially for dynamically changing terrains. This will enable novel footstep planning and robot localization in the environment. Analytic and (self-supervised and reinforcement) learning methods will leverage multi-modal sensing to allow quadrupedal robots mimic the way that animals plan footsteps when learning to walk. The developed methods will be validated experimentally on several full-size quadrupedal robots, in academic and industrial real-world use cases, for tasks such as inspection, patrolling, and maintenance. RoboHike will work towards the next-generation autonomous robotic systems in construction fields, oil&gas sites, or damaged sites after a man-made/natural disaster, where efficient navigation is required, and rough/rocky terrain, industrial stairs, pipes, and narrow passages may exist. The vision is to endow quadrupeds with environment cognition for the benefit of the public in autonomizing manual labor of hard or dangerous tasks. The impact is expected to be high in the national and industrial sectors for automated inspection, monitoring, maintenance, and disaster innervations, where terrain is arduous and the requirement for timely intervention is paramount. We intend to construct publicly shared benchmark datasets on challenging trails, bringing in this way the robotics community several steps forward in robot locomotion by enabling robots to work on challenging grounds.