SBIR Phase I: Learning from Demonstration for Customer-Grade Robots

这个小企业创新研究（SBIR）第一阶段项目的更广泛影响是提高当前机器人自主参与家庭环境，识别物体并以有目的的方式操纵它们的能力。所提出的方法将这种能力添加到广泛的家用和娱乐机器人中。用户可以教机器人广泛的自主行为，而不是以前安装的基线软件包。提出的技术从根本上降低了教机器人复杂行为的门槛，而无需复杂的编程或高质量的演示。这使得一类新的应用超越了简单的重复工业活动。这个小企业创新研究（SBIR）第一阶段项目利用了一个基于深度神经网络的计算机视觉系统，该系统可以创建机器人世界观的低维表示。在这种潜在编码中，使用优化标准（例如训练损失）的组合来训练机器人的行为，以确保目标任务的泛化，平稳运行，从意外错误中恢复，以及在问题的替代解决方案之间进行选择的能力（例如，避开左边或右边的障碍物）。这项技术是新颖的：目前通过演示系统进行的学习通常会重现与已演示的相同的轨迹。这可能在工厂车间有用，但在家庭环境中不起作用。相比之下，所提出的模型学习一种可推广的行为，该行为使用演示来获得关于如何解决特定操作问题的线索，并且可以识别并从自己的错误中学习。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

The broader impact of this Small Business Innovation Research (SBIR) Phase I project is to advance the capability of current robots to autonomously engage with the household environment, recognize objects, and manipulate them in a purposeful way. The proposed methodology adds this capability to a wide range of household and entertainment robots. A user can teach a robot a wide range of autonomous behaviors out of reach of the previously installed baseline software packages. The proposed technology radically lowers the threshold for teaching a robot sophisticated behaviors without complex programming nor high-quality demonstrations. This enables a class of new applications beyond the simple repetition of industrial activities. This Small Business Innovation Research (SBIR) Phase I project leverages a deep neural network-based computer vision system that creates a lower-dimensional representation of the world-view of the robot. On this latent encoding, the robot behavior is trained using a combination of optimization criteria (e.g. training loss) that ensures the generalization of the target task, smooth operation, recovery from accidental mistakes, and ability to choose between alternative solutions to the problem (e.g. avoid an obstacle to the left or to the right). The technology is novel: current learning by demonstration systems usually reproduce an identical trajectory to what has been demonstrated. This might be useful on a factory floor but would not work in a home environment. In contrast, the proposed model learns a generalizable behavior that uses demonstrations to obtain clues as to how to solve a particular manipulation problem and can both identify and learn from its own mistakes.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.