Mutually Stabilized Correction in Physical Demonstration

Abstract:

The goals of this project include the development of real-time control for human-machine co-control of highly dynamic and potentially dangerous systems. The work focuses on formalizing the automated assessment of trust, primarily focusing on the degree to which a computer should trust a human operator. We are focusing on two experimental tasks: a) a crane operation task, where control in automation settings requires a highly skilled operator, and b) rehabilitation and training tasks. However, these tasks are only examples; all the software is being implemented in the Robot Operating System (ROS) in ways that can be deployed for a very general set of tasks. The project has focused on experimental development in the health science portions of the grant and experiments using crane control. The project has several real operator-in-the-loop CPS systems working and theoretical developments can be tested on actual hardware with people involved in operator-in-the-loop control. The first project is on automated sensory augmentation for stroke survivors using ROS and real-time control. In this project we use tactors (small motors used in mobile phones for vibrate mode) to provide feedback to subjects engaged in virtual balance tasks (in this case balancing a simulated inverted pendulum). We find that if we provide state feedback, performance changes very little. However, if we provide control feedback by actually synthesizing a controller in real-time that solves the balance problem, people not only perform better when using the tactors, they perform better later after the tactors have been taken away. The key need is to automate the measures of these increases in performance. The second project is on crane control, including operator-in-the-loop real-time experimental control using variational integrators that have been fully integrated in ROS. The project has now demonstrated tight control of a crane system in the laboratory using model-predictive control based on variational integrators. The most interesting aspect of this work is that the numerical representations of the real-time system play a major role in its performance. Using a standard numerical method, both the estimation and control fail at data rates greater than 500 Hz even in simulation. Using the variational integrator-based methods, the estimation and control functions both in simulation and experiment at 20 Hz, 10 Hz, and even 6 Hz (below the Nyquist cutoff frequency). The tight control of the crane allows less experienced operators to direct its motion, largely because the inexperience operators have a more nave cognitive model of the dynamics. Again, automating the measurement of how much control authority the computer should assign the operator becomes a critical part of the work because the cognitive model is constantly changing. We have additionally found that when one uses variational integrators, Extended Kalman Filters and particle filters both improve dramatically and become nearly identical to each other (at least in the context of the crane examples). We have now completed the total integration of variational integrators and their associated filtering and trajectory optimization techniques, in ROS. The third project has been the development of a simulated of a double-link inverted spherical pendulum experiment for trust-based control in the context of 3D balance tasks using a balance board. We now have our first preliminary data showing that machine learning techniques can adapt trust in real-time based on data.