Context-Aware Learning from Demonstration: Using Camera Data to Support the Synergistic Control of a Multi-Joint Prosthetic Arm

摘要

Ahstract- Muscle synergies in humans are context-dependent-they are based on the integration of vision, sensorimotor information and proprioception. In particular, visual information plays a significant role in the execution of goal-directed grasping movements. Based on a desired motor task, a limb is directed to the correct spatial location and the posture of the hand reflects the size, shape and orientation of the grasped object. Such contextual synergies are largely absent from modern prosthetic robots. In this work, we therefore introduce a new algorithmic contribution to support the context-aware, synergistic control of multiple degrees-of-freedom of an upper-limb prosthesis. In our previous work, we showcased an actor-critic reinforcement learning method that allowed someone with an amputation to use their non-amputated arm to teach their prosthetic arm how to move through a range of coordinated motions and grasp patterns. We here extend this approach to include visual information that could potentially help achieve context-dependent movement. To study the integration of visual context into coordinated grasping, we recorded computer vision information, myoelectic signals, inertial measurements, and positional information during a subject's training a robotic arm. Our approach was evaluated via prediction learning, wherein our algorithm was tasked with accurately distinguishing between three different muscle synergies involving similar myoelectric signals based on visual context from a robot-mounted camera. These preliminary results suggest that even simple visual data can help a learning system disentangle synergies that would be indistinguishable based solely on motor and myoelectric signals recorded from the human user and their robotic arm. We therefore suggest that integrating learned, vision-contingent predictions about movement synergies into a prosthetic control system could potentially allow systems to better adapt to diverse situations of daily-life prosthesis use.