NRI: FND: Efficient algorithms for safety guiding mobile robots through spaces populated by humans and mobile intelligent machines and robots

这项研究工作的目标是为最近旨在促进智能(自主或半自主)机器和机器人在我们日常生活的不同方面安全和顺利地整合的努力做出贡献。人类和不同类型的机器人将不得不在包括密集的城市场所(例如繁忙的十字路口)和大型工业设施的工厂车间在内的共享空间中共存和共同工作。对于机器人和人类的和谐共生，需要两者都能有效地避免可能造成严重破坏、重大经济损失、伤害甚至生命损失的冲突和物理碰撞。这项研究工作的关键挑战之一与这样一个事实有关，即当机器人在其他机器人和人类居住的繁忙空间中导航时，它们将不得不在不确定的情况下实时做出决定。特别是，移动机器人在人口稠密空间安全导航的能力不仅取决于它们对附近其他移动机器人或人类的下落的了解，还取决于后者的意图(即它们计划朝哪个方向移动)，这可能很难预测。这项研究将创建新的算法方法，使机器人能够同时1)实时推断附近人类和机器人最有可能的未来运动模式，2)安全地引导它们到达目的地，同时避免与附近的机器人和人类发生碰撞。这一研究工作有望为多智能体环境中自主机器人的分散意图感知局部运动规划创建可扩展的算法。该方法明确考虑了参与冲突事件的智能体的移动特征和形状，以及不确定性对其决策机制的影响，这是由于(1)智能体的感知/感知限制，以及(2)感兴趣的智能体(自我-智能体)对其附近智能体关于其未来运动的意图缺乏了解。我们方法的主干之一是意图识别算法，该算法寻求计算与冲突情况中涉及的每个代理的预测目标目的地相关的概率分布的密度函数的近似。该算法依赖于一类被称为核密度估计算法的非参数统计方法，其计算量和复杂度明显小于其他依赖于部分可观测马尔可夫决策处理(POMDP)的解的方法，这可能是一项非常复杂的任务。随后，我们构造椭球管，它以一定的概率包含预期的轨迹，这些轨迹将把自我代理附近的所有代理转移到他们计划的目标目的地。接下来，我们将局部运动规划问题归结为一个低维的凸优化问题，只有当智能体未来轨迹的预测发生重大变化时，该问题的解才会更新。所提出的方法明智地描述了碰撞的高风险区域，从而允许代理即使在密集拥挤的空间中也能够规划无碰撞的轨迹。这与基于可达性的方法形成对比，后者经常对碰撞避免问题中的可行性问题给出错误的否定答案，从而宣布后一个问题是不可行的，尽管存在可行的无碰撞轨迹。还提出了一系列相互交织的研究和教育活动，以促进本科生和代表性不足的学生参与现实世界中的机器人问题。该奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

The goal of this research effort is to contribute to recent efforts aimed at promoting the safe and smooth integration of intelligent (autonomous or semi-autonomous) machines and robots in different aspects of our everyday life. Humans and different types of robots will have to co-exist and work together in shared spaces including dense urban places (e.g., busy crossroads) and the factory floors of big industrial facilities. For the harmonious symbiosis of robots and humans, it is necessary that they can both effectively avoid conflicts and physical collisions that may cause serious damage, significant economic losses, injuries, or even loss of life. One of the key challenges of this research effort has to do with the fact that robots will have to make decisions in real-time under uncertainty when they are navigating through busy spaces populated by other robots and humans. In particular, the ability of mobile robots to safely navigate in densely populated spaces hinges upon their knowledge of not only the whereabouts of the other mobile robots or humans in their vicinity but also the intentions of the latter (i.e., which directions they plan to move) which may be difficult to predict. The proposed research will create new algorithmic methods that will allow robots to simultaneously 1) infer the most likely future motion patterns of nearby humans and robots in real-time and 2) safely guide them to their destination while avoiding collisions with nearby robots and humans.This research effort is expected to lead to the creation of scalable algorithms for decentralized intention-aware local motion planning for autonomous robots in multi-agent environments. The proposed approach explicitly accounts for the mobility characteristics and the shape of the agents involved in a conflict event as well as the effects of uncertainty on their decision making mechanisms due to (1) sensing / perception limitations of the agents, and (2) lack of knowledge by the agent of interest (ego-agent) of the intentions of its nearby agents regarding their future motion. One of the backbones of our approach is an intention identification algorithm that seeks to compute an approximation of the density function of the probability distribution associated with the projected goal destination of each agent involved in a conflict situation. The proposed algorithm relies on a class of non-parametric statistical methods known as kernel density estimation algorithms whose computational footprint and complexity are significantly smaller than those of other approaches that rely on the solution of partially observable Markov decision processed (POMDPs), which can be a very complex task. Subsequently, we construct ellipsoidal tubes that contain with a certain probability the anticipated trajectories that will transfer all the agents near the ego-agent to their projected goal destinations. Next, we reduce the local motion planning problem to a low-dimensional convex optimization problem whose solution will be updated only when significant changes in the predictions of the agents future trajectories have taken place. The proposed approach judiciously characterizes the areas of high risk for collisions, thus allowing the agents to plan collision-free trajectories even in densely crowded spaces. This is in contrast with reachability-based approaches which often give false negative answers to the question of feasibility in a collision avoidance problem, thus declaring the latter problem to be infeasible despite the existence of feasible collision-free trajectories. An array of interweaved research and educational activities that will promote the participation of undergraduate and underrepresented students in real world problems of robotics are also proposed.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.