Vestibular Bipeds

[Music] this is a vestibular biped what's a vestibular biped it's a creature from an artificial life simulation that uses a genetic algorithm to evolve a population of creatures to achieve bipedal locomotion it's described in a paper I recently wrote that will be published in the 2023 artificial life conference proceedings okay to start with vestibular refers to the sense of balance and equilibrium in this case the concept is simplified as a linear object that can detect when it is tilted away from the direction of gravity and bipad refers to the fact that these Critter can walk on two leges how do they learn to walk well the individual creatures themselves don't learn to walk rather a population of genes learns to walk through dark Ian Evolution as thousands of individuals are born and die their genes pass through many generations and undergo natural selection I'll come back to that in a moment but first I want to talk about simulated physics so one of the first animation tricks I Learned was how to make a bouncing ball not long after that I learned how to attach two balls with a spring if you connect three balls with three Springs you get a flexible triangle add one more ball and three more more Springs and you get a tetraedron and here's where things start to get interesting using only balls and springs a simple three-dimensional object can be simulated it's not as precise as a rigid body but it's one way to approximate certain physical systems by adding a ball over the face of a tetrahedron or tet for short and attaching it with three Springs you get a new shape made of two tets and this is how I designed the creatures in these experiments they come in several body types and all of them are made of tetrahedra including one tet two 3 four and five tets since walking on two feet is a tricky Balancing Act I began my experiments with something simpler tripedal Locomotion walking on three feet here are some examples of evolved tripeds once I got that working I could then move on to the vestibular walking now every time a new ball and three Springs are added the creature becomes more wobbly and difficult to control I like to imagine that pretty much anything could be simulated using Springs and point masses that is balls with infinitely small radi but only up to a point for the sake of experimentation and Discovery sometimes simpler is better I decided to reduce the components of a bipedal creature such as a chicken to its Bare Essentials a body a head a tail and two feet this is a fet creature made of eight point masses and 18 Springs it's hardly a chicken but it has plenty of wobble and I'm very interested in putting that wobble to good use using the power of evolution many robot folks and artificial life researchers are familiar with Valentino bronberg Vehicles simple car-like robotss that are great at illustrating some basic concepts there are two sensors in front that detect light and two actuators in back that cause Wheels to turn sensors can be connected to actuators Crossing these connections causes a vehicle to move towards the light those connections could be crafted into neural nets for some interesting results the wiring in between is analogous to a brain which takes information from the senses and makes decisions on how to control the actuators these springy creatures have sensors that are part of their bodies instead of sensing light they sense tilt in their own bodies the vular sense of balance as complex intricate and tightly coordinated with eye movement and muscle reactions I started to wonder how I could apply the principle of balance to simulated creatures so I reduced it to the simplest possible representation a line segment that is a spring in the body of the creature whose Direction in global space is compared to the direction of a point in the environment for balance that point is the Zenith an imaginary point above that is opposite the direction of gravity so what about actuators well instead of applying torque on a wheel as in a brenberg vehicle these actuators adjust the lengths of Springs like muscles the 2D computer simulation soda play from the early 2000s was based on this very principle Springs that change length according to sign functions there are now many other cool examples of this technique being used for simulated Locomotion imagine an equilateral triangle where two edges change length in a periodic fashion the top vertex moves up and down if the sine waves are out of phase with each other the vertex moves in a different way similarly a tetrahedron with three edges changing lengths with different phes can Trace out all different kinds of paths depending on those phase shifts at first I was going to try to simulate the effects of opposing muscle systems but that would have made things a bit complicated so I decided to go even more abstract you could think of the bottom points of each of these examples as end affectors that make contact with the horizontal surface and yes some octopuses can walk every creature has a heartbeat and that heartbeat determines the frequency of a master sine wave that is the basis for all oscillating spring lengths if all the Springs changed according to this master sine wave the creature would pulse but the amplitudes and phases of those sine waves are genetically adjusted for each individual spring and here's what happens if the phases and amplitudes are shifted randomly painful chaos the purpose of the genetic algorithm is to find needles in the hay stack motions that actually allow the creature to walk the Tilt sensor continually adjusts the phases and amplitudes for each spring the genetics of a creature determine exactly how the signals from the Tilt sensor are used to adjust the springs in order to keep the creature from falling over when you lose your balance your muscles automatically adjust in ways that are too complex and unconscious for you to be aware of likewise it would be nearly impossible for you to design the parameters for balance by hand for that you need natural selection the genetic algorithm calls the worst creatures allowing the better ones to reproduce and generate Offspring it's designed by Evolution after many generations of reproduction and death the genes of the population Converge on a good solution here's an illustration of all the sine waves of an 18 spring creature this is the base motor system of the creature the central pattern generator it's not constrained by any symmetry but since the tetrahedral bodies have bilateral symmetry the central pattern generator tends to go for symmetrical Solutions but not always and here's an important thing to point out a lot of chaotic wobbling goes on because of the jell likee bodies of these creatures the genetic algorithm takes all this wobbling in stride and comes up with ways to work with it all the secondary motions and physical Dynamics are incorporated into the solution this method is consistent with the claim that walking is more a matter of physical passive Dynamics than direct top down neural control here's the central pattern generator with the wobble included as a red line This wobble is exploited by natural selection some of these creatures can walk on an undulating surface because they are continually adjusting their Central pattern generator in response to their sense organs check this out using the same basic idea of a tilt sensor I can add a Target sensor instead of using the Zenith for sensing the upward Direction it detects the direction to a moving Target in the environment each sensor is wired up to Every Spring in the creature and affects the amplitudes and phases of their sine waves all I have to do is wire them up and the genetic algorithm does the rest a creature that immediately tumbles over into a quivering Heap gets a low Fitness and is unlikely to mate with another creature to produce Offspring and those creatures that are relatively more fit that is able to walk closer to the Target and not fall over are more likely to reproduce this genetic algorithm was designed especially for this experiment there are three components to Fitness how close the creature is to the Target on average over its lifespan how how long the creature remains upright before falling onto a nonf foot point and how energy efficient it is Energy Efficiency is measured as the inverse of average spring Force it's used as a fitness penalty the details of all of this are explained in the paper Valentino brenberg wrote a paper for the aife proceedings in 1989 called some types of movements in contrast to simp aquatic animals having similar symmetry to his Vehicles larger animals with articulated appendages have motor control systems that must deal with many degrees of freedom they pose fascinating problems of interpretation because we do not know in what terms the states and changes of the motor system are represented in the brain and he ends with this it is my contention that the brain deals with movements in a naive way so so I took this as permission to build a naive model to discover how a motor control system might evolve in a toy simulation where the fitness criteria are abstracted away from the gnarly details of bodily mechanics [Music]

Channel: Jeffrey Ventrella

Views: 59,865

Keywords:

Id: o79ndlTgeMw

Length: 11min 7sec (667 seconds)

Published: Sun Apr 28 2024