Soft RoboticsEvolve the body and brain

of `soft’ robots.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Robots are composed of a setof voxels, rather than a collectionof rigid 3D shapes.

Some of the voxels are hard orsoft;

some of the voxels are `muscles’:they change in volume.

HyperNEAT is used to evolvethe bodies.

The robots have no brain:their ‘muscles’ pulse at a regularfrequency.

‘muscle’ type 1: change volume

‘fat’: soft, deformed by neighboring voxels

muscle type 2: change volume in antiphase to mt1

bone: rigid, is not deformed by neighboring voxels

Voxel = three-dimensional pixel

Soft RoboticsCan create more complex

robots than `rigid’ ones

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Why little progress in 19 years?

Evolution has few options ifasked to put together a few shapes;

Much more design freedom forlarge number of voxels withdifferent material properties

Soft RoboticsWhy use HyperNEAT?

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Robots are composed of a setof voxels, rather than a collectionof rigid 3D shapes.

Voxels have different materialproperties (e.g. hard/soft)

Would like to have regular 3Dpatterns of different kinds ofvoxels throughout the robot.

X Y

Greyscale of pixelxy

X Y Z

Deposit plastic droplet (y/n)

Soft RoboticsEvolve the body and brain

of `soft’ robots.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Take a CPPN.

Feed in the x, y, z position of eachpixel within a cube. Also d, itsposition from the cube’s center.

One binary output value:pixel present/absent

four continuous output values:pixel type 1, … pixel type 4If pixel present, place pixel typewith maximum value.

If multiple patches, take the oneclosest to the center.

max( )

Soft RoboticsPhysical soft robots.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Physical robot is composed oftwo voxel types: red and white.

Robot placed in a pressurechamber.

Q: How does the robot move?

Soft RoboticsSoft robots

of variable resolution.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Genotype = CPPN(CompositionalPattern-ProducingNetwork)

Can produce multiple phenotypes(robots) at different resolutions.

Simply use smaller voxels andrequery the CPPN.

Re-query CPPN at

higher resolutio

n

(below is from lecture 23)

Soft RoboticsCompare HyperNEAT

to a direct encoding

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Q: How would you define agenotype to directly encode a10x10x10=103 voxel soft robotcomposed of four differentvoxel types? How many `genes’would there be?

Soft RoboticsEvolve the body and brain

of `soft’ robots.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Generative encodingproduces regular patternsof the same voxel type.

“In the direct encoding, each voxel works independently from—and often at odds with—its neighboring voxels, preventing coordinated behaviors.” (section 4.1)

Soft Robotics

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

FitnessFn1: displacement (d)

Fitness functions 2 through 4:add a penalty term of the form

F = d * ( 1 - penalty/maxPenalty )

FF2: d * (1–usedVoxels/1000)

FF3: d * (1 – connsBetVoxels/ maxConnsBetVoxels )

FF4: d * (1–usedActVoxels/1000 )

How different fitness functionsaffect evolution.

Soft RoboticsHow different fitness functions

affect the robot phenotypes.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

FitnessFn1: displacement (d)

Fitness functions 2 through 4:add a penalty term of the form

F = d * ( 1 - penalty/maxPenalty )

FF2: d * (1–usedVoxels/1000)

FF3: d * (1 – connsBetVoxels/ maxConnsBetVoxels )

FF4: d * (1–usedActVoxels/1000 )

Soft Robotics

Soft RoboticsEvolution

of biologically-similarphenotypes.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

Soft RoboticsDo the different voxel types

matter?

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

35 evolutionary runs……with all four voxel types.

35 evolutionary runs… …with red, blue and green voxels.

35 evolutionary runs… …with just red and blue voxels.

35 evolutionary runs……with just red voxels.

‘muscle’ type 1: change volume

‘fat’: soft, deformed by neighboring voxels

muscle type 2: change volume in antiphase to mt1

bone: rigid, is not deformed by neighboring voxels

Soft RoboticsGetting beyond

locomotion overflat ground.

N Cheney, R MacCurdy, J Clune, H Lipson. Unshackling Evolution:

Evolving Soft Robots with Multiple Materials and Powerful Generative

Encoding. GECCO 2013.

How to get to object manipulation?

Start by ‘grabbing’ obstacleson the ground and pushingor pulling yourself forward..