Sensing without a Sensor
Proprioception is the ability of the human body to sense movement, action and location. It is what enables us to dribble a basketball without looking at it. It is the ability to move without thinking. “Sixth sense” in some sense (no pun intended).
The easiest and most direct way to enable proprioception in robots is to install sensors. A force sensor to measure point forces, a temperature sensor to measure heat, a flex sensor to measure the degree of bending, a torque sensor to measure torque generated, so on and so forth. This means that to sense 10 elements, 10 individual sensors are needed. This is not usually a problem for most robotic applications, but as the size of the robot gets smaller, the amount of space and volume available gets smaller too. So if 10 sensors are needed but only 5 can be installed, that then becomes a problem.
Then I thought, why not give robots a “brain” then? You know how our human brain can derive solutions when given information, or lack thereof? Like how other senses heighten in a deaf or blind person. Essentially, I want to program a brain that can derive information in the absence of some. The keyword here is some.
In the soft actuators that I have designed so far, pressure sensors are absolutely necessary. They are needed to prevent the actuators from exploding, quite literally. My idea is hence, to use pressure readings to derive multimodal information. To me, this concept is intuitive.
Say you are doing some stretching exercises and bending your trunk. How will you know that you are bending to the left? It’s probably because you feel the stretch in the trunk muscles on your right and a compression on the ones on your left. And if you are bending to the right, the opposite happens.
So, how will a soft actuator know it is bending left or right? Is it then not based on the internal pressures in its chamber(s)?
The method I used to process these pressure readings is through machine learning, a LSTM network to be precise. (You probably guessed where I was going with this when I mentioned that I wanted to program a brain earlier.) It is basically to throw a bunch of pressure readings into the algorithm and let it recognize patterns to predict its bending state. Eventually, I was able to derive multimodal information of both bending angle and force output through my methodology but I will not go into that much detail here. If you are interested, the full article can found here.
Overall, I am very satisfied with what I was able to accomplish. With my method, the number of sensors needed is reduced and in the event of sensor failure, some redundancy is available. I also added the skill of programming a machine learning algorithm to my arsenals and I could not be happier.
