The aim of this project is to improve the stability of a Passive Dynamic Walker (PDW) by only controlling the position of an added torso. The desired angle of torso is determined by using a neuroevolutionary algorithm that tries to maximize the distance walked by robot. The “brain” of this robot is a fully connected feedforward neural network with two hidden layers. The evolutionary algorithm used here is a simple Genetic Algorithm (GA) that initially starts with random weights for the feedforward network and then favors individuals in a generation that were able to walk for a more distance. This setting enabled the walker to walk on wider range of surface slopes (+0.01 to -0.2 radians) in comparison to a full passive walker without a torso (-0.005 and -0.1 radians).

The walking gait of the robot in zero slope angle is seen in the following video:

The following video shows the robot walking down the slope of -0.15 radians:

These two examples clearly show that the devised controller has extended the range of slopes that the robot can walk on stably. An interesting outcome of this algorithm is how it uses the strategies familiar to humans to keep itself in balance. This can be observed especially as the slope reaches its two extremes. For example, the robot has intuitively learned that it needs to lean back when walking down hill and lean forward in walking on straight line.

The next objective of this work is to decrease the robot’s sensitivity to initial conditions. The measure we used for this objective is the area of the basin of attraction of the robot’s walking cycle. The following figure shows how the basin of attraction for the robot changes after the addition of torso and its associated controller.

Calculations show that after applying the control strategy the basin of attraction grows by approximately 58 percent. Therefore the second objective of this project is also realized.

The results of this project has been presented in the 26th Iranian Conference on Electrical Engineering (ICEE 2019). The full paper is available at the link below.