Latest / Elon Musk Podcast / AI robots and drones shepherding desert sheep

Transcript

AI robots and drones shepherding desert sheep · Elon Musk Podcast

0:00
So researchers Parikh Shitmani and Kita Shukla and Andrew Hess
0:05
over at the University of Nevada, Reno are building this
0:08
4000 LB autonomous robot. Right named Robohydra.
0:12
Exactly Robohydra. It carries 1100 lbs of water and
0:16
it guides flocks of sheep across the desert using facial
0:19
recognition. AI track every single animal's
0:21
health. It is.
0:23
I mean, it's wild. You are basically looking at a
0:25
collision between highly advanced robotics and one of
0:29
humanity's absolute oldest professions, right?
0:32
This is precision livestock farming in its most literal
0:35
physical sense. We are taking the mechanics of a
0:37
giant roving water tank and, you know, employing it as a
0:41
mechanical shepherd to actively manage animal welfare out in
0:45
incredibly harsh, unforgiving conditions.
0:47
Yeah, and we have giant stack of sources on the table for you
0:49
today, covering everything from drone engineering schematics to
0:53
agricultural research papers on desert ecology.
0:56
Our mission for you listening is to explore what happens when you
1:00
remove human beings from the management of natural
1:02
landscapes. So the question guiding us today
1:04
is, if we replace the human shepherd, the sheepdog and the
1:08
static watering hole with robotics and artificial
1:11
intelligence, how does that fundamentally alter the biology
1:16
of the flock and the ecology of the earth they walk on?
1:18
Well, to understand the ground robotics, we actually need to
1:21
look up first. I mean, we have to start with
1:23
the sky, OK? Edward Behre Cloth and Doctor
1:25
Sebastian Hahn created a system called Drone Hand and this is an
1:30
autonomous drone system actively managing livestock on these
1:34
super remote farms that stretch across 1,000,000 acres in
1:37
Australia. Wow. 1,000,000 acres.
1:39
Yeah. And the engineering hurdle they
1:42
had to solve right out of the gate is that those remote range
1:44
lands possess absolutely 0 Internet access.
1:47
I mean there is no cell service out there.
1:49
You are not connecting to cloud computing to process your data
1:53
which. Means the drone basically has to
1:54
do all the thinking by itself. Exactly.
1:56
These drones have to operate entirely offline.
1:59
All the machine learning, all the image processing, it happens
2:02
right there on the drone's internal hardware as it
2:05
physically flies over the paddock.
2:07
So they are visually identifying water levels in troughs and they
2:11
are identifying cast animals. Yeah, and for those of you who
2:14
are familiar with sheep anatomy, a cast animal is when a sheep
2:17
gets physically stuck on its back or its side, right?
2:20
Because of the shape of their bodies, their center of gravity,
2:22
and just the heavyweight of a full sleece, they simply cannot
2:26
roll over once they get stuck. No, they care.
2:28
And their digestive system continues to produce gas.
2:32
That gas bloats the stomach, compresses their lungs, and it
2:36
becomes fatal very quickly. Yeah, it is a bad situation.
2:40
O Having an offline camera spot that from the air is obviously a
2:43
massive survival tool, but I'm looking at this drone setup and
2:48
I essentially see a low orbit satellite for sheep.
2:52
Basically, yeah. So how does putting an offline
2:54
camera in the sky change the actual business of farming?
2:57
Well, it limits an enormous financial burden for the
3:00
operators, specifically their reliance on helicopter
3:03
mustering. Oh, gathering thousands of
3:05
animals across 1,000,000 acres requires hiring teams of
3:09
helicopters to physically push the flock toward the pens.
3:13
That costs a single operation upwards of $120,000 for just a
3:18
couple of weeks of flying. That is crazy.
3:20
And the failure rate is incredibly high.
3:22
Traditional helicopter mustering only has a 70% success rate.
3:26
Right, because you have a jet engine screaming over a prey
3:28
animal. Exactly.
3:30
The noise, the sudden shadows and the violent downdraft from
3:33
the rotor wash. They all just cause the animals
3:35
to panic. They scatter, they hide under
3:37
heavy brush and 30% of the flock just gets left behind in the
3:41
wild. So drone hand changes that
3:43
dynamic entirely. An autonomous drone applies
3:46
steady, quiet, consistent pressure.
3:49
It doesn't terrify the animals. And that stealth approach
3:52
increases the capture rate and totally eliminates that huge
3:56
helicopter bill. That makes a lot of sense.
3:58
Furthermore, they're applying this exact same computer vision
4:01
technology to fixed camera systems.
4:04
They are testing these with meat companies like JBS in high
4:08
density feedlots, where cameras monitor the pens constantly to
4:12
detect early signs of disease through behavioral observation.
4:16
Yeah, they are looking for an animal that removes itself from
4:19
the group, right? Because if you think about how
4:21
prey animals survive, their entire defense strategy is based
4:25
on herd mentality. You stay close to the middle of
4:28
the group, so the predator eats the guy on the edge.
4:30
Exactly. So a single sheep willingly
4:32
isolating itself in the corner of a feed lot is a glaring
4:36
behavioral anomaly. The algorithm flags that
4:39
isolation, alerting the human managers to a sick animal long
4:43
before physical symptoms like coughing or lethargy ever
4:45
appear. So watching from the air is
4:47
highly efficient in those open scenarios, but the physical
4:50
environment dictates severe limits on what airborne sensors
4:53
can actually see. The sky is not a perfect vantage
4:56
point. Wait, hold on back up.
4:57
I want to challenge that based on the thermal technology we see
4:59
in these source papers. If a drone has a high end
5:02
thermal camera, shouldn't a warm blooded mammal glowing against a
5:06
cold high altitude desert be the easiest thing in the world for a
5:10
sensor to spot? Well, you would naturally assume
5:13
so, but you have to account for the physical behavior of the
5:16
terrain itself. Marcus Blum and his research
5:19
team studied the limitations of unmanned aerial vehicles, or UA
5:23
VS, in mountain sheep habitats, and they encountered A
5:27
phenomenon known as thermal clutter.
5:29
Thermal clutter. Yeah, desert substrates,
5:32
specifically rocks, boulders and granite cliffs, possess very
5:36
high thermal inertia. Oh, think of like a cast iron
5:38
skillet. You take it off the stove, but
5:40
it stays burning out for a long time.
5:42
Precisely. These rocks absorb solar
5:44
radiation all day and retain that heat well into the evening,
5:47
right? So when the sun bakes those
5:49
rocks, the temperature contrast between the 98° animal and the
5:53
background environment completely disappears.
5:56
The sensor looks down and just sees a solid white wall of heat.
5:59
Oh wow. Yeah, the thermal contrast drops
6:01
to 0 and the sheep becomes totally invisible to the
6:04
machine. And on top of that, you have the
6:06
biology of the animal fighting the sensor, right?
6:09
Desert bighorn sheep have evolved specifically to avoid
6:12
severe heat. Their natural biological defense
6:15
mechanism against high ambient temperatures is to seek heavy
6:19
shade deep inside caves or under massive rock out crops.
6:23
Yeah, exactly. And they also do this to escape
6:26
aerial predators like Golden Eagles.
6:29
So when a bighorn sheep retreats into a cave, it renders thermal
6:32
infrared sensors totally useless because the drone can't see
6:36
through a rock ceiling. Right and standard optical
6:39
cameras fail because the animal is hidden in deep shadows.
6:42
Yeah, so the $1,000,000 technology is completely blinded
6:45
by the animals natural instinct to hide.
6:47
Exactly. Plus you also have severe human
6:50
constraints governing the sky. Pilots operating these drones
6:53
struggle constantly with visual line of sight regulations.
6:56
By law, the operator often has to maintain direct visual
6:59
contact with the aircraft. Which is tough.
7:02
Yeah, if you are flying in a Canyon environment with sudden
7:04
topographic changes, maintaining that visual contact from a
7:08
stationary position is nearly impossible.
7:10
Yeah, and plus there is the very real physical threat of crashing
7:15
an expensive piece of equipment into a sudden Cliff face because
7:19
the terrain changes faster than the pilot can react.
7:21
Exactly so this. Limits the entire premise of
7:24
aerial monitoring. We simply cannot rely solely on
7:28
the sky to understand an ecosystem.
7:31
The failure of aerial sensors in rugged rocky terrain opens up
7:35
the absolute necessity for ground level intervention.
7:38
Which, you know, brings us directly back to Rubble Hydra in
7:41
Nevada. Because aerial monitoring has
7:44
massive blind spots, the technology must physically roll
7:47
alongside the animals. Right, so Robohydra operates as
7:50
this massive mobile resource hub, and let's just look at the
7:53
actual engineering mechanics of the robot for a second.
7:55
It utilizes a completely closed water tank in a high desert
8:00
environment with low humidity and highwind, open water
8:03
evaporates almost instantly. Yeah, you'd lose it all.
8:05
Exactly. So to prevent losing the water
8:07
to the air, the tank feeds into a push activated bowl.
8:10
OK. The water is only exposed to the
8:12
environment when the sheet physically presses its nose
8:15
against the mechanism to drink. Oh, that's smart.
8:17
And as that happens, highly sensitive flow meters measure
8:21
the exact volume of water intake for that specific drink.
8:26
And simultaneously, cameras mounted right above the bowl
8:29
capture detailed facial images of the animal.
8:32
And they are doing this entirely without ear tags.
8:34
Right, because anyone who works with livestock knows ear tags
8:38
are just a nightmare. They really are.
8:40
To get snagged on thick brush and ripped out, they cause
8:42
infections, and they require humans to physically catch and
8:46
restrain the animal just to read the plastic number.
8:49
Yeah, it's. A huge hassle.
8:50
So by using facial recognition artificial intelligence, the
8:53
system maps the exact geometry of the sheep's face.
8:56
It measures the distance between the eyes, the pigment patterns
9:00
on the nose, the specific structure of the jaw.
9:03
It identifies the individual animal, records exactly how many
9:06
milliliters of water it just consumed, logs its exterior
9:10
surface temperature, and ties all of that data to a permanent
9:13
digital profile. And the onboard computer
9:16
analyzes all of this collected data to make its most critical
9:19
decision, which is when and where to move Next.
9:22
The robot evaluates the hydration levels of the flock
9:24
against the surrounding forage conditions, and then it simply
9:27
drives away. And the sheep, driven by their
9:30
deep biological need for water, just follow the robot across the
9:34
landscape. Which changes the entire grazing
9:37
structure of the environment. I mean it completely eliminates
9:40
what ecologists call sacrifice zones.
9:43
Historically, ranchers pipe water from a well to a static
9:47
heavy concrete through sheep are tethered to that singular water
9:52
source. Yeah, they won't leave.
9:54
Right, they will not walk 5 miles away to eat fresh grass.
9:57
They will stay close to the through and over graze the
10:00
immediate surrounding area until it is completely barren dirt.
10:03
Exactly. And that creates A sacrifice
10:05
zone where nothing grows. Meanwhile, remote, highly
10:10
nutritious forage just over the next mountain Ridge goes
10:13
completely untouched because it is simply too far from the
10:15
water. Yeah, so by putting the water
10:18
tank on wheels and giving it an artificial brain, the robot
10:21
dictates the grazing pattern of the entire flock.
10:23
And we can connect this localized control to the broader
10:26
ecology of the Great Basin. Tracy Shane's work focuses on
10:30
virtual fencing and targeted grazing.
10:32
Virtual fencing uses GPS collars on the animals to emit audio
10:36
cues and mild physical stimuli to keep the herd within a
10:39
specific invisible digital boundary.
10:42
Right. By steering the animals to very
10:44
specific geographic areas, ranchers force the herd to eat
10:48
invasive species, particularly cheat grass.
10:51
Yeah, and we really need to explain cheat grass for you
10:53
listening because it is destroying the American West.
10:56
It really is. Native bunch grasses grow in
10:58
distinct, isolated clumps with bare soil between them.
11:02
That bare dirt acts as a natural firebreak, stopping a wildfire
11:06
from spreading across the ground.
11:08
But cheat grass is this invasive weed that fills in all those
11:12
bare patches, creating a continuous, highly flammable
11:15
fuel bed that basically connects the native plants like a fuse.
11:19
Yeah, it's highly flammable. But there is a brief biological
11:22
window, the green up phase in early spring, where cheat grass
11:25
is actually palatable and highly nutritious for sheep.
11:28
Exactly. The goal is to hit that sheet
11:30
grass heavily with grazing pressure right at that specific
11:33
phase, before it dries out and turns to kindling.
11:35
And to prove this targeted grazing actually works,
11:39
researchers use drones paired with terrestrial laser scanners.
11:44
The drones fly above to provide the overhead canopy view, but
11:48
the terrestrial lasers on the ground shoot sideways.
11:51
They penetrate the dense outer leaves of the shrub to
11:53
accurately measure the three-dimensional volume of the
11:56
woody plant biomass hidden underneath.
11:59
And this combined sensor data confirmed that forcing sheep to
12:03
graze these specific targeted areas reduces wildfire fuel from
12:07
£400 per acre down to just £100 per acre.
12:11
That is wild. This opens up a reality where
12:14
sheep are utilized as highly precise biological wildfire
12:19
prevention tools. They really are.
12:20
The technology allows us to place the exact right number of
12:23
mouths on the exact right hillside at the exact right
12:25
moment. Yeah.
12:26
So just to reset our pacing here, we are literally steering
12:29
sheep with robotic water tanks and invisible fences to eat the
12:32
kindling before the forest has a chance to burn down.
12:34
That's. Exactly what's happening and
12:36
tracking every single bite of cheat grass and every single
12:39
drop of consumed water generates A staggering amount of data.
12:43
That data is then used to physically alter the future of
12:45
the animal itself through selective breeding.
12:47
OK, how so? Well, Andrew Hess conducts
12:50
genetic research using the Rafter 7 Merino flock, and this
12:53
specific flock is a crossbreed, right?
12:55
It combines the Rambouye, which is a rugged, Hardy breed built
12:59
to survive brutal climates and sparse forage, with the
13:03
Australian Merino, which produces incredibly fine,
13:06
premium quality wool. And the biological mechanism
13:10
they use to track the success of these genetics is just
13:12
fascinating. Yeah, it is.
13:13
They track seasonal changes in the physical diameter of a
13:16
single wool fiber, right? Because wool growth is a
13:19
continuous biological process, if an animal experiences severe
13:23
heat stress or a lack of nutrition during a drought, its
13:26
body automatically diverts energy away from producing wool
13:30
and redirects that energy toward keeping its vital internal
13:33
organs functioning. And this physical survival
13:36
response causes the actual diameter of the growing wool
13:39
fiber to shrink. By measuring a single staple of
13:42
wool from the base to the tip, Hess creates a physical
13:46
historical record of the animals stress and nutrition over the
13:49
entire year. Exactly like reading the rings
13:52
of a tree to understand historical weather patterns.
13:54
And this specific biological data opens up precision
13:58
genetics, right? Researchers use this information
14:00
to build highly detailed Pangeno maps.
14:03
Instead of relying on a single reference genome, A Pangenom
14:07
captures the genetic diversity across the entire population.
14:11
You are not just looking at one ideal sheep, you are mapping the
14:14
very variations of thousands of sheep to find the hidden traits
14:18
that allow certain individuals to thrive.
14:21
They use this map to selectively breed sheep for thermal
14:24
tolerance, for feed efficiency, and for natural parasite
14:27
resistance. They are purposely adapting the
14:29
DNA of the flock over generations to survive an
14:32
increasingly harsh arid climate. OK, I completely understand the
14:36
science of mapping the genetics, but looking through our sources
14:39
I have to push back on this entire approach.
14:41
Oh, really? Yeah, I see all these research
14:43
papers on automated indoor wool farms and it makes me wonder,
14:47
why spend generations painstakingly tweaking DNA just
14:51
so an animal can barely survive in a brutal burning desert when
14:55
we possess the engineering capacity to completely control
14:58
the environment? Well, that is the other side of
15:00
it. Right, because inside these high
15:02
tech barns, sheep live in perfect climate controlled
15:05
environments. There is zero heat stress.
15:08
There are no mountain lions. There is never a lack of water.
15:12
Autonomous robots drive down the aisles and deliver precise
15:15
nutritional diets directly to the pens based on the sheep's
15:17
exact weight. When it is time for harvesting,
15:21
automated shearing robots use advanced sensors to safely
15:24
remove the fleece without stressing or cutting the animal.
15:27
Then optical scanning technology categorizes the raw wool by
15:30
color and fineness instantly on a conveyor belt.
15:33
Yeah, it is incredibly efficient.
15:35
The entire facility feeds into a circular economy where recycled
15:39
wool reduces waste and the building runs on renewable
15:41
energy. We can build a perfectly safe,
15:43
perfectly optimized world for the animal indoors right now.
15:46
You can. And what this split opens up for
15:49
you are two radically different philosophies regarding the
15:52
future of agriculture and Land Management.
15:54
Yeah. Your indoor model uses
15:56
artificial intelligence and robotics to optimize the
16:00
immediate environment to protect the animal from the harsh
16:02
realities of nature. Right.
16:04
The Nevada model uses artificial intelligence and robotics to
16:07
optimize the animal's genetics to protect the outdoor
16:10
environment from catastrophic fire and ecological collapse.
16:14
Oh wow. Yeah, 1 strategy separates the
16:16
biological system from the earth entirely, and the other strategy
16:20
integrates the animal more deeply into the dirt to heal the
16:22
landscape. So essentially we were replacing
16:25
static physical infrastructure with mobile intelligence, using
16:29
autonomous systems to monitor everything from individual water
16:33
intake to this specific geometry of a sheep's face.
16:36
Exactly. And this allows us to treat a
16:37
flock of thousands as individual patients, optimizing their
16:41
biological health while simultaneously deploying them as
16:44
tools to manage wildfire risks across the open range.
16:47
Right. And you know, we covered how
16:48
humans use artificial intelligence to direct the
16:51
flock. But consider this, as these
16:53
systems grow more sophisticated, what happens when the artificial
16:56
intelligence controlling the movements of the robotic
16:59
shepherd begins communicating directly with the artificial
17:02
intelligence predicting long term climate models?
17:04
Oh man. Right.
17:06
At what point does the machine entirely bypass the human ranch
17:09
manager moving animals across the earth based on ecological
17:13
logic we cannot even erceive. That is wild.
17:16
Well, if you're not subscribed yet, take a second and hit
17:18
follow on. Whatever a you're using, it
17:20
helps us Kee making this. We appreciate you being here.