Shared Control For Dynamic Systems Between Humans and Robots
Researchposted by Kaylen Sanders, ODSC September 14, 2018 Kaylen Sanders, ODSC
Engineering and computer science researchers at Northwestern University devised a model that forsakes user-only control in favor of an optimized shared control paradigm.
As machines become further integrated in our daily lives, we will naturally see a rise in human-machine systems that use human input while also relying on automated decision-making. In these situations, control of the machine necessitates a balancing act, raising questions about what degree of human involvement should be permitted and when is it most useful. For example, a system with major safety implications might benefit from tighter restrictions on human control. Alternately, an instance like piloting an aircraft should delegate low-level tasks to automation and maintain human input from the pilot, who is best able to respond to unprecedented environmental variables. In any case, the ideal human-machine system is one that can dynamically shift control between human and robotic management as needed and as fit for the system as a whole.
The researchers approached this task as an optimization problem. Rather than creating a rigid model fit to a specific user and system, they wanted to design a model that could generalize to various systems and adapt to the behavior of individual users. To do this, they utilized the Koopman operator, “an infinite-dimensional linear operator that can capture all relevant information about any nonlinear dynamical system.” First, the researchers would perform data collection to gain knowledge about a particular user’s interaction with the system, and then they set out to establish a learned framework for shared control by achieving outer-loop stabilization. Including the learned models in the shared control paradigm led to a significant improvement in performance of the system, evidenced by the difference in ergodicity between shared control versus user-only control experiments.
With an adaptive model that incorporates automated intervention tailored to a user’s skill level, human-machine systems can advance beyond rigid determinism and surmount performance bottlenecks. For future work, the researchers will consider using higher-order, nonlinear basis functions to represent the Koopman operator rather than the previously used linear basis function. The goal is that this will facilitate better understanding of respective learned models. Another potential area of ongoing investigation involves analyzing how data collection informs the computation of the Koopman operator and its resulting influence. Find out more here.