Latest / Elon Musk Podcast / AI UPDATE: Massive Tesla Optimus Robot China News

Transcript

AI UPDATE: Massive Tesla Optimus Robot China News · Elon Musk Podcast

0:00
A major electric vehicle manufacturer is halting
0:02
production of two of its flagship car models just to
0:06
convert an entire automotive factory into a dedicated
0:09
humanoid robot manufacturing facility.
0:11
Yeah, I mean, the sheer physical scale of that transition is
0:14
staggering. We're looking at a fundamental
0:16
shift from building consumer vehicles to, you know, churning
0:21
out human sized machines designed to integrate directly
0:25
into the physical economy. Right.
0:26
And it brings up so many questions about the engineering
0:29
breakthroughs making this hardware possible, plus the
0:32
software architecture that actually allows a machine to
0:34
learn how to manipulate objects, the massive manufacturing
0:37
targets, and of course, the fierce competition already
0:40
deploying robots in the real world.
0:42
Exactly. It gives us a really clearview
0:43
of where physical labor is heading.
0:46
So how exactly does a machine transition from a clumsy, rigid
0:50
piece of hardware into like a fully autonomous entity capable
0:54
of doing your household chores? And will you actually want one
0:57
living in your house? Well, the answer to that starts
1:00
at the very end of the robotic arm.
1:03
The newest generation of this general purpose robot features
1:06
hands with 22° of freedom. Oh wow, 22.
1:10
Yeah, and they're driven by 50 custom actuators packed entirely
1:14
into the robot's forearms. Which directly mirrors human bio
1:17
mechanics right? Like if you look at your own
1:19
arms right now and move your fingers, you do not actually
1:21
have muscles in your fingers. Right, exactly.
1:23
Your forearm muscles are doing all the heavy lifting.
1:25
Yeah, they just pull on long tendons that cross your wrist to
1:28
operate the fingers, and the engineers did exactly the same
1:31
thing here. By keeping the bulky, heavy
1:34
electric motors up in the forearms and using a
1:37
sophisticated high tension cable system that functions just like
1:41
human tendons, the robot's fingers stay incredibly slim and
1:45
agile. And that agility allows for an
1:47
extreme level of precision. The internal sensors achieve a
1:51
control that allows for a force sensitivity of 0.1 Newtons.
1:56
Wait back up. Why is force sensitivity that
1:58
specific so important? Because that precise measurement
2:02
allows the machine to physically feel the difference between
2:06
grasping A rigid metal tool like a hammer and holding a fragile
2:10
object like an egg or a piece of soft fruit.
2:12
So it won't crush it. Right.
2:14
It knows exactly how much pressure to apply without
2:16
dropping the object or crushing it.
2:18
It adjusts that pressure in three milliseconds based on the
2:21
resistance it feels. That response time is actually
2:24
faster than human reflexes. It is.
2:26
Yeah, and the positioning accuracy of the joints
2:29
themselves is 0.05°. I read that every single gear
2:35
bearing and coupling inside the hand is machined to aerospace
2:39
tolerances. Which is wild, but high
2:41
precision electric motors generate massive amounts of heat
2:44
when operating continuously. So to counteract that, the
2:47
engineers integrated micro cooling channels directly into
2:50
the metal housing. Wait, really?
2:52
Yeah. Liquid coolant flows
2:53
continuously through internal passages thinner than a human
2:56
hair. That keeps the operating
2:57
temperatures perfectly stable even when the machine is lifting
3:00
heavy loads or gripping an object for hours at a time.
3:03
So that extreme physical sensitivity and mechanical
3:05
precision totally eliminates the need for specialized robotic
3:08
attachments. You do not need a machine with a
3:10
barcode scanner or an electric drill physically bolted onto its
3:14
wrist. Exactly.
3:16
Opens up the ability for the hardware to use unmodified human
3:19
tools directly off the shelf. It picks up a standard kitchen
3:22
knife, operates a regular computer keyboard or uses a
3:26
manual hand saw. It completely changes how
3:28
automation integrates into existing human workspaces
3:32
because, well, the hardware adapts to our tools.
3:35
We don't have to rebuild our factories, offices and homes
3:38
just to accommodate the machine. Right.
3:40
And processing all that complex movement requires taking in
3:43
staggering amounts of sensory data continuously.
3:46
We are talking about the machine processing 1.2 terabytes of data
3:50
every single hour. Yeah, to put that volume of data
3:52
into perspective, that is the processing equivalent of
3:55
watching hundreds of high definition movies
3:58
simultaneously. All of this information is
4:00
pushed through a single unified neural network running on a
4:04
custom artificial intelligence chip mounted inside the unit.
4:08
It's kind of like catching a baseball while talking to a
4:10
friend. You are not consciously pausing
4:13
your speech to run complex physics calculations on the
4:16
ball's trajectory and some isolated part of your brain.
4:19
Right, exactly. Your brain just fuses your
4:21
vision, your sense of touch, and your audio perception into one
4:24
fluid action. And that is exactly what this
4:27
custom chip is doing. Engineers call it multimodal
4:30
fusion. It combines vision, touch, audio
4:33
and physical balance into one continuous stream of decision
4:36
making. The cameras map the geometry of
4:39
the room, the microphones process ambient noise and voice
4:42
commands, the internal gyroscopes maintain balance
4:45
against gravity, and the four sensors in the fingertips
4:48
measure the grip strength. All of the exact same
4:51
microsecond. Which is just an incredible
4:53
amount of input. Yeah, and the custom silicon
4:56
ship delivers ultra low latency, meaning the data flow never
4:59
bottlenecks. The machine doesn't have to
5:01
pause, calculate a specific trajectory, and then execute a
5:05
movement. It just reacts fluidly to
5:07
whatever is happening in its environment.
5:09
And that neural architecture completely removes the need for
5:12
manual line by line programming. Traditional factory hardware
5:17
requires a human engineer to code every single specific
5:21
movement. If you want a traditional arm to
5:23
pick up a box, you program the exact spatial coordinates for
5:27
the hand to reach. Right, but this unified neural
5:29
network allows the hardware to learn autonomously through
5:32
digital simulation and by observing first person video of
5:36
humans actually performing physical tasks.
5:38
Just by watching video, yeah. And while one physical unit
5:41
practices a task in the real world, thousands of virtual
5:44
copies of the robot are training inside a massive digital
5:47
simulation environment. Oh wow, they practice infinite
5:50
edge cases, failure scenarios, and complex multi step tasks
5:54
that would be frankly too dangerous or time consuming to
5:56
test physically. A single night of simulation
5:59
equals months or even years of real world physical practice.
6:03
The physical unit then downloads that accumulated perfected
6:05
knowledge. So if the physical machine
6:07
reaches for an object and fails to grasp it correctly, it
6:10
analyzes its own mistake, adjusts its neural pathways, and
6:14
tries a completely new physical approach.
6:16
It creates infinite adaptability for unstructured environments.
6:20
And Speaking of unstructured environments, the newest
6:22
iteration of the machine is fully water resistant with
6:25
advanced polymer shielding 70% of its body to protect the
6:29
internal circuits from moisture, dust and corrosion.
6:32
It also utilizes autonomous charging.
6:35
It uses onboard computer vision to perfectly align itself with a
6:38
charging dock without any human intervention.
6:41
It leverages the exact same high density battery cell technology
6:45
originally designed for electric vehicles, giving it an
6:48
incredibly long operational life on a single charge.
6:51
Hold on, we have to look at the practical reality of operating
6:54
this equipment. A multi $1000 piece of highly
6:57
advanced hardware is essentially A paperweight if a spilled glass
7:01
of water or a sudden rain shower destroys the main circuitry.
7:04
I mean true, but the technical specifications of the movement
7:07
are amazing. The processing power is massive,
7:10
the fluid movement mimics biology perfectly, and the
7:13
motors operate at whisper quiet decibel levels.
7:17
It's a huge stick up, sure. But waterproofing and self
7:20
charging aren't luxury features to make the spec sheet look
7:22
impressive. They are absolute necessities
7:25
for survival. If this technology is going to
7:28
succeed outside of a pristine climate controlled laboratory,
7:31
it has to survive the brutal physical environment.
7:34
That is a fairpoint that protective shielding changes the
7:37
unit from a fragile laboratory experiment into a reliable
7:40
appliance. It enables continuous operation
7:43
in highly unpredictable settings.
7:44
Think about outdoor gardens, wet concrete sidewalks, or
7:47
incredibly messy commercial kitchens.
7:49
Right. It limits the amount of human
7:51
supervision required to zero. You do not follow the machine
7:54
around with an umbrella, and you don't manually plug it into the
7:57
wall when the battery gets low. It monitors its own internal
8:00
power consumption and mathematically calculates when
8:03
it needs to return to base to dock itself.
8:06
And you know, while that highly anticipated general purpose
8:09
hardware is currently only deployed internally, testing
8:12
inside its own manufacturing facilities, A competitor named
8:15
Agility Robotics is already operating its digit units in
8:19
massive commercial warehouses around the country.
8:22
Digit takes a very different approach to the mechanical
8:24
engineering. It uses a unique 4 bar linkage
8:27
leg design. Instead of human knees that bend
8:30
forward, requiring constant muscle tension and battery power
8:33
just to stand upright, A4 bar linkage allows the hardware to
8:37
lock its legs passively into a standing position.
8:40
So it spends almost 0 battery power while standing perfectly
8:43
still. Exactly.
8:45
Instead of focusing on highly complex human like fingers, it
8:49
relies on specialized purpose built grippers designed
8:52
specifically to handle plastic totes and cardboard boxes.
8:55
But having human like hands is obviously superior for future
8:58
applications, right? If you want a machine to exist
9:01
in spaces build exclusively for humans, it needs to be able to
9:04
manipulate the physical world exactly like a human does.
9:07
Specialized clamp grippers seem like a total dead end if the
9:10
ultimate goal is general utility.
9:11
I actually disagree on that. One digit has a near perfect
9:15
success rate at moving boxes precisely because it ignores
9:19
human anatomy completely. It focuses strictly on logistics
9:22
and material handling. The grippers on the end of its
9:25
arms do not need to fold laundry, thread a needle, or
9:28
cook a meal. They just need to secure a
9:30
plastic tote and move it from point A to point B.
9:33
OK, that makes sense. Yeah, because it does not carry
9:36
the immense computational and mechanical burden of operating a
9:39
22° of freedom hand digit is available right now.
9:43
They use a robotics as a service model where commercial
9:45
warehouses lease machines by the hour, solving real manual labor
9:49
shortages today while the really human like hands are still being
9:52
refined behind closed doors in the lab.
9:55
Well, then you have a company called Unitary that produces the
9:57
G1 humanoid model. The G1 is incredibly agile.
10:02
It's capable of running fast, recovering balance seamlessly
10:05
after being pushed or kicked, and performing complex physical
10:08
acrobatics like backflips. And the base model costs only
10:12
$16,000. But severe security
10:15
vulnerabilities were discovered in the unitary operating
10:17
software. Independent security researchers
10:20
found critical flaws, including unauthorized data exfiltration.
10:24
Yeah. That's a huge issue.
10:25
The hardware sensors, like the onboard cameras and the
10:28
microphones, were actively transmitting environmental data
10:31
out to external servers without the operator having any
10:34
knowledge of it. They also discovered backdoor
10:37
access points hidden in the code and a wormable Bluetooth
10:40
vulnerability. And justice to clarify, a
10:43
warmable vulnerability means the malicious code sreads
10:46
automatically from unit to unit, right?
10:48
Exactly. That specific flaw allows an
10:50
outside attacker to take full remote control of the machine.
10:54
The wormable aspect means the infection spreads automatically
10:58
from one compromise unit to any other vulnerable unit within
11:02
Bluetooth range, without any human interaction required.
11:05
Yeah, you could have a single compromise unit enter a
11:08
facility, and within minutes it transmits the payload to every
11:11
other robot in the building. An attacker could view live
11:14
camera feeds, listen to internal microphones and remotely
11:17
manipulate the physical movements of an entire fleet of
11:20
machines. Which drastically limits the
11:22
environments where the G1 can be deployed safely.
11:25
Security flaws of that magnitude shut the hardware completely out
11:28
of sensitive government, military, or corporate
11:31
facilities. No secure operation can risk a
11:33
mobile camera and microphone platform transmitting
11:36
unauthorized data or being hijacked by someone sitting with
11:39
a laptop in the parking lot. Right.
11:41
But even with those severe software vulnerabilities, the
11:44
physical existence of the G1 proves that highly capable,
11:48
affordable hardware is already a reality.
11:52
The mechanical barriers to entry to build a walking balancing
11:55
robot are falling rapidly. Which brings us to the ultimate
11:59
manufacturing target for the general purpose robot to reach
12:02
an output of 10 million humanoid units annually at a massive
12:06
dedicated facility currently under construction in Texas.
12:09
Hold on 10 million units a year. Yeah, that volume is roughly 20
12:13
times the current global production of all industrial
12:16
robots combined. That is massive.
12:18
They plan to achieve that huge volume by utilizing the unboxed
12:22
manufacturing method. If you look at the traditional
12:25
automotive assembly line, the main chassis moves slowly down a
12:29
single linear conveyor belt from start to finish, with workers
12:32
bolting parts on one by one. Right, the classic assembly
12:35
line. But the unboxed method
12:38
completely abandons that linear process.
12:41
Different lens and sub assemblies are built
12:43
simultaneously on completely parallel lines.
12:46
The arms, the internal torso, skeleton and the legs are all
12:49
assembled independently by automated systems and then
12:53
brought together at a central hub for final integration.
12:55
So drastically speeds up the entire manufacturing process.
12:58
Exactly. It removes massive bottlenecks
13:01
and reduces the physical footprint of the factory.
13:03
And the scale of that production ties directly back to you, the
13:06
consumer. The target retail price for the
13:08
unit is $20,000. For less than the cost of a new
13:12
compact car, You could own a machine that handles your
13:14
tedious daily chores, or a small business could own an automated
13:18
worker that operates around the clock without demanding overtime
13:21
pay. And this level of mass
13:23
production scales directly into the concept of a von Neumann
13:26
machine. John von Neumann was a
13:28
mathematician who proposed the theoretical idea of self
13:31
replicating machines. You build a machine that has the
13:34
physical capability and the programmed intelligence to build
13:37
an exact physical copy of itself.
13:39
Like robots building robots. Exactly.
13:42
The machines mine the raw materials from the earth, refine
13:45
those materials into usable metal and plastic components,
13:48
and assemble those components into new machines.
13:51
It creates an exponential growth curve in manufacturing capacity.
13:55
1 Machine builds 22 Build 44 build 8.
13:58
It completely shifts the global economic bottleneck.
14:01
If you look at human history, economic output has always been
14:05
strictly limited by human labor availability.
14:07
You can only harvest as much food, build as many houses, or
14:11
manufacture as many cars as your human workforce can physically
14:14
handle. Right, because humans need to
14:16
sleep, they get injured, and there is a finite number of them
14:19
available for physical labor in any given geographic area.
14:22
But if you introduce millions of self replicating autonomous
14:26
workers into the economy, that traditional bottleneck
14:29
completely disappears. The.
14:31
Absolute limit to global economic growth shifts away from
14:35
human labor availability. It become entirely dependent on
14:38
the physical availability of raw materials and energy to power
14:41
the machines. Which fundamentally alters the
14:43
cost of goods and services worldwide.
14:46
Think about the physical cost of an apple at the grocery store.
14:50
You are paying for the human labor to plant the tree, the
14:53
manual labor to pick the apple, the labor to drive the transport
14:57
truck to the store, and the labor to stock the physical
14:59
shelf. Yeah, that adds up.
15:01
When the cost of the physical labor required to perform all
15:03
those mechanical steps drops toward absolute 0, the
15:07
underlying cost structure of the entire global economy changes.
15:11
So we are moving away from rigid single purpose hardware bolted
15:15
to factory floors into an era of highly dexterous self learning
15:20
equipment that adapts to our existing spaces rather than
15:23
forcing us to adapt our environments to theirs.
15:27
The hardware has the precision to handle fragile objects, the
15:29
software has the neural architecture to learn
15:31
autonomously from its physical mistakes, and the manufacturing
15:35
infrastructure is currently being built to produce them by
15:38
the millions. If a commercially available
15:40
machine can observe, learn, and perfectly replicate any physical
15:43
task you do, what becomes the unique value of human physical
15:47
skill? If you're not subscribed yet,
15:48
take a second and hit follow on whatever app you're using.
15:51
It helps us keep making this. We appreciate you being here.