Distributed Sensing, Learning and Control in an Assistive Manipulator

One of the grand challenges for the robotics community is to create robots that operate robustly in real- world scenarios. Most current robots are limited to factories, laboratories or similar controlled settings. This contrasts with the seeming ease with which insects, animals and humans handle uncertainty, dynamic events and complexity. Assistive robots are for example being envisioned for aiding elderly and disabled persons in their homes. A key skill for these robots will be to operate in, and physically manipulate, daily life environments. However, it is unclear how to achieve this while complying with the safety and reliability requirements of such devices. Distributed Adaptive Control (DAC) is an example of a biologically inspired architecture for control and adaptation, where the lowest unit is the reflex. We here explore recent work on extending this idea to shared control of assistive robot manipulators. That is, where sensing, learning and control are distributed throughout the system, and across both user and robot. We show that such a distributed approach can reduce the need for central information processing, exact internal representations, and "global" approaches to learning in the robot. The reduced algorithmic complexity can help increase the robustness and usability of the system on real-world tasks.