Action-based Object Representations for Planning and Control
Planning and control are essential components of any robot system, yet they have long been studied in isolation. This, we contend, is due to the lack of a common representational framework. We are exploring a holistic approach to planning robot behavior, using previously acquired skills to represent control knowledge (and objects) directly, and to use this background knowledge to build plans in the space of control actions. We are also developing a learning framework in which the agent builds a functional model of interactions with the world.
Truly dexterous mobile manipulators will adaptively exploit properties of their bodies and the environment to achieve high-performance control. We are exploring ways that redundancy and dynamic interaction between mobility and manipulation resources can be used to increase efficiency, robustness, and versatility with the uBot-5 and uBot-6. Experiments with the uBot-5 include whole-body pushing, learning a simple knuckle-walking gait for traversing rough terrain, and learning to exploit rapid arm motions to recover from perturbations while balancing.
Embedded Health Care in the Activity of Daily Living
We study the efficacy of humanoid-mediated health care in an embedded setting, especially targeting stroke survivors. In traditional approaches, the hardware design of specially designed exoskeleton robots is confined to a particular therapy activity, limiting its applicability. In our approach, we employ general purpose robots and program various therapy activities that are prescribed by experienced therapists. In this approach, developing a) robots, b) therapy activities, c) robot behavior programming methodology, and d) human-robot interaction can be addressed independently in a synergetic way.
Contemporary stroke rehabilitation studies mostly address single type of rehabilitation separately, such as physical therapy only or speech therapy only. Hence, it is not known how different types of therapy activities can affect the results or progress of each other. We study the effects of inter/multidisciplinary rehabilitation, such as physical and speech therapy activities through humanoids. This will help us understand how multiple different types of therapy activities may be provided together and how they can be scheduled to maximize the effects synergistically.
An autonomous system's sensitivity to risk (i.e., performance variance) will, in general, be a function of its current operating context (e.g., the presence of dangerous obstacles, battery charge, etc.). We are developing model-free policy learning algorithms that support on-the-fly switching between different risk profiles. Experiments with the uBot-5 include learning arm motion strategies to stabilize after large impacts and learning to exploit object dynamics for heavy lifting.
Autonomous Skill Acquisition and Learning Skills from Demonstration
An important type of cumulative learning is the acquisition of procedural knowledge in the form of skills. Allowing a robot to abstract away from low-level motor control and plan and learn at a higher level should improve its problem solving abilities, and in turn create further opportunities for learning. In our experiments with the uBot-5, we showed that both expert demonstrations and learned sequences innate controllers can serve as example solutions from which a component skills may be extracted. The resulting skills improve the robot’s ability to learn to solve related tasks.