NRI: Collaborative Research: Learning Deep Sensorimotor Policies for Shared Autonomy

辅助机器人有潜力改变上肢残疾人的生活，帮助他们进行基本的日常活动，例如操纵物体和喂食。然而，辅助机器人的人类控制面临着巨大的挑战。机械臂的高维性意味着类似操纵杆的界面不自然，难以直观地使用，并且直接远程操作产生的运动通常缓慢、不精确，并且灵活性受到严重限制。这项研究通过开发共享自主学习算法来解决这些挑战，其中机器人可以预测用户的意图并提供一定程度的辅助自主以确保流畅和成功的运动。这项研究还将为未来的研究铺平道路，这些研究可以从远程操作引导到机器人的完全自主。该研究提出了一种利用深度学习和强化学习技术来实现共享自治的分层、多阶段方法。该系统首先使用深度逆强化学习来快速确定用户的高级目标，例如用户是否想要抓住特定物体或操作设备，从原始感官输入。该目标推理层为较低的控制层提供目标，该控制层由深度神经网络控制策略组成，可以直接处理有关环境和用户的原始感官输入以做出决策。这些策略选择低级控制来满足高级目标，同时最大限度地减少与用户命令的分歧。这些算法将在安装在轮椅上的机器人手臂上进行部署和测试，有可能帮助上肢残疾的用户进行日常生活活动。

Assistive robots have the potential to transform the lives of persons with upper extremity disabilities, by helping them perform basic daily activities, such as manipulating objects and feeding. However, human control of assistive robots presents substantial challenges. The high dimensionality of robotic arms means that joystick-like interfaces are unnatural hard to use intuitively, and motions resulting from direct teleoperation are often slow, imprecise, and severely limited in their dexterity. This research address these challenges by developing learning algorithms for shared autonomy, where the robot anticipates the user's intent and provides a degree of assistive autonomy to ensure fluid and successful motions. This research will also pave the way for future research that can bootstrap from teleoperation and build towards full robot autonomy. The research proposes a hierarchical and multi-phased approach to shared autonomy, using techniques from deep learning and reinforcement learning. The system begins by using deep inverse reinforcement learning to quickly ascertain the user's high-level goal, such as whether the user wants to grasp a particular object or operate an appliance, from raw sensory inputs. This goal inference layer supplies objectives to the lower control layer, which consists of deep neural network control policies that can directly process raw sensory input about the environment and the user to make decisions. These policies choose low-level controls to satisfy the high-level objective while minimizing disagreement with the user's commands. The algorithms will be deployed and tested on a wheelchair-mounted robot arm with the potential to assist users with upper extremity disabilities to perform activities of daily living.