Description of the Neural Network Thruster Control Video - 1994

Edward Wilson, Intellization

The project and technology are described in the associated papers and thesis. This work was performed at the Stanford Aerospace Robotics Laboratory, under the direction of Prof. Steve Rock and Prof. Robert Cannon. This is a thruster-propelled space-robot simulator supported on an air-bearing, allowing it to simulate the zero-g, drag-free motions of space. The video demonstrates fully autonomous recovery to multiple abrupt de-stabilizing thruster failures.

The robot nominally has eight thrusters symmetrically located as shown on the monitor in the lower right corner (a mechanically implemented picture-in-picture). This monitor displays the robot's understanding of the state of its thrusters. Unknown to the robot, 6 of the 8 thrusters have been mechanically failed by partial plugging (50% thrust), full plugging (0% thrust), partial (45 degree) and full bending (90 degree). The two 90-degree bent thrusters are destabilizing since they change the sign of applied torque.

Initially, the robot is near-motionless within its control deadband, but when it eventually fires one of the 90-degree-bent thrusters, the instability is triggered and it begins to spin out of control. The on-board accelerometers and angular rate sensor are used by the thruster identification system to detect and identify this change in thruster properties. A directly-calculated stabilizing neural network is calculated and immediately implemented. Stability is recovered within 4 seconds.

As time progresses, more failed thrusters are detected and identified (due to their on/off nature, they cannot be detected unless they are commanded to fire). During this time the robot is regulating to the initial desired state, but exciting unknown failure modes as possible while maintaining position and attitude. The monitor shows the progress in identifying failed thrusters.

Concurrently with this FDI process, a neural-network training process is continually optimizing a thruster controller to accommodate the newly discovered thruster properties. The neural network implemented in the control loop is periodically updated depending on its performance on off-line testing. The training was performed on an early 1990s era SUN workstation running MATLAB. About 20 seconds of training time is not shown in the video (at the 40 second mark of the video).

After one minute (t = 40 seconds in the video), sufficient control capability has been regained, and the robot does a final 3-axis maneuver to demonstrate the control capability that has been restored in the presence of these significant thruster failures.