The Econmist - USA (2021-11-06)

72 Science & technology The Economist November 6th 2021

builds up precise layers of material under

robotic control, and greatly reduces waste.

“Using much less cement is a very impor­

tant part of the answer,” he adds, especially

as cement production looks otherwise set

to double over the next 20 years.

Additives  can  also  make  concrete  last

longer  and  reduce  the  need  for  mainte­

nance. At the University of Michigan, Vic­

tor Li and his colleagues use synthetic and

natural fibres, along with CO 2 injection, to

produce a bendable concrete they call En­

gineered  Cementitious  Composite  (ecc).

The internal structure of this material was

inspired by nacre, a flexible material com­

monly  called  “mother  of  pearl”  that  coats

the  insides  of  the  shells  of  molluscs  such

as abalone and oysters.

Adding such flexibility to concrete lets

bridges  and  roads  cope  more  easily  with

heavy traffic, and improves the earthquake

resistance  of  tall  buildings.  eccdevelops

only tiny surface cracks when it ages. Dr Li

says  it  is  thus  better  at  keeping  water  out

and  preventing  corrosion  of  reinforcing

steel bars inside. Such corrosion can cause

reinforced­concrete structures to crumble

within a few years of their construction—

sometimes resulting in their collapse.

To zero and beyond

Substitution of materials could go still fur­

ther.  Solidia,  a  firm  in  New  Jersey,  makes

cement containing calcium silicates with a

higher ratio of silica to calcium oxide than

the  standard  “Portland”  variety.  This  has

two  consequences.  One  is  that  Solidia’s

process  requires  less  heat  (and  therefore

less fossil fuel) than conventional calcina­

tion,  and  so  releases  less  CO 2 in  the  first

place.  The  other  is  that,  when  mixed  into

concrete,  Solidia’s  silica­rich  silicates  can

be cured more rapidly than regular cement

by  using  captured  CO 2 instead  of  water.

Solidia  is  working  on  applications  for  its

cement  with  one  of  its  investors,  Lafarge­

Holcim, a Swiss building­supplies giant.

Taking all these developments into ac­

count,  how  green  could  concrete  get?  Dr

Fennell says it would be reasonably easy to

reduce  the  industry’s  CO 2 emissions  to

around 80% of present levels per tonne of

concrete  produced  by  better  energy  use

and  the  modification  of  materials.  But

companies could really pull the stops out if

they moved to kilns largely or entirely po­

wered by biomass, such as wood. The car­

bon in this would, until recently, have been

CO 2 in the air. If, after being turned back in­

to  that  gas  by  being  burned  in  the  kiln,  it

was stored away and not released, the con­

sequence,  as  new  trees  grew  to  replace

thoseconsumed,wouldbeanetflowof

carbonoutoftheatmosphere.

Thissortofsystem,calledbioenergy

withcarboncaptureandstorage(beccs), is

onewayclimatemodellersimaginepro­

vidingthe“negativeemissions”neededfor

net­zeroornet­negativeemissionstargets.

beccs­basedelectricitygenerationisoften

talkedof,butbeccsmightactuallybebet­

tersuitedtocement­making—becauseina

carbon­consciousworldtheCO 2 ­capturing

equipmentwillalreadybethere,dealing

withresultsofcalcination.Andif thathap­

pened,oneofthepariahsofglobalwarm­

ingmightthusredeemitselfbyhelpingal­

leviatethedamagebeingdonetotheplan­

et,andsoleavebehinda legacyasimpres­

siveinitswayasthatoftheRomans.n

Unimals

Balls, sticks and

the Baldwin effect

I

t mightsoundobviousthatifyou want

to improve a robot’s software, you should

improve its software. Agrim Gupta of Stan­

ford University, however, begs to differ. He

thinks you can also improve a robot’s soft­

ware  by  improving  its  hardware—that  is,

by  letting  the  hardware  adapt  itself  to  the

software’s capabilities. 

As they describe in Nature Communica-

tions, he and his colleagues have devised a

way  of  testing  this  idea.  In  doing  so,  they

have  brought  to  robotics  the  principles  of

evolution  by  natural  selection.  They  have

also  cast  the  spotlight  on  an  evolutionary

idea  that  dates  from  the  1890s,  but  which

has hitherto proved hard to demonstrate.

There  is  a  wrinkle.  The  team’s  robots,

which  they  dub  “unimals”,  are  not  things

of metal and plastic. Rather, they are soft­

ware  entities  that  interact  with  a  virtual

environment  in  the  way  that  metal­and­

plastic  devices  might  interact  with  a  real

one.  Unimals  are  pretty  simple,  having

spheres  for  heads  and  cylinders  for  limbs

(see  picture).  The  environments  through

which  they  roamed  were  also  simple,  and

came in three varieties: flat arenas, arenas

filled with hills, steps and rubble, and ones

that  had  the  complexities  of  the  second

sort, but with added props like cubes that

needed to be moved around.

To begin with, the unimals were given a

variety  of  randomly  assigned  shapes,  but

with  identical  software  running  each  of

them. That software was a piece of artificial

intelligence called a deep evolutionary re­

inforcement learning algorithm, or derl. 

Newly created unimals started in a vir­

tual boot camp, in which the derllearned

enough  about  the  world  to  face  the  chal­

lenges to come. They were then entered in­

to tournaments. In groups of four, Dr Gup­

ta put them through tests of agility, stabil­

ity and ability to manipulate objects. Each

group’s  winner  was  allowed  to  “breed”  by

spawning  a  daughter  with  one  mutation

(an extra limb for stability, perhaps, or ex­

tra rotation in a joint, for flexibility). This

daughter was substituted for the oldest un­

imal in the pool, assigned to a new group of

four, and the process repeated. 

Unimals were withdrawn from the fray

after  ten  generations  of  evolution,  and  Dr

Gupta  reckons  about  4,000  varieties  of

them  underwent  training.  The  team  were

surprised  by  the  diversity  of  shapes  that

evolved.  Some  had  arms  as  well  as  legs.

Others  had  only  legs.  There  were  bipeds,

tripeds and quadrupeds. Some moved like

lizards. Others resembled an octopus walk­

ing  on  land.  Crucially,  though,  the  re­

searchers  found  that  the  most  successful

unimals learned tasks in half the time that

their  oldest  ancestors  had  taken,  and  that

those which evolved in the toughest arenas

were the most successful of all. 

In this evolution of unimals’ morpholo­

gy to promote the ability to learn, Dr Gupta

sees  a  version  of  something  called  the

Baldwin  effect.  In  1896  James  Baldwin,  an

American psychologist, argued that minds

evolve  to  make  optimal  use  of  the  mor­

phologies  of  the  bodies  they  find  them­

selves  in.  What  Dr  Gupta  has  shown,

though in software rather than in the real,

biological world, is that the obverse can al­

so  be  true—changes  in  body  morphology

can  optimise  the  way  minds  (or,  at  least,

derls) work. Even though he held the soft­

ware  constant  from  generation  to  genera­

tion,  it  became  more  efficient  at  learning

as the unimals’ bodies evolved. 

Whether  that  discovery  can  be  turned

to account in the way robots are developed

remains  to  be  seen.Butitis  certainly,  in

the  jargon  beloved  ofsomebusinessfolk,

an out­of­the­box idea.n

A novel way to optimise robots

CorrectionAn eagle-eyed reader has spotted that

the picture illustrating one of last week’s Science

and technology stories, “No sex please, we’re

condors”, was of an Andean condor, not one of the

Californian variety. Sorry.