Kind: captions
Language: en
You know, for decades, robots have been
these onetrick ponies, right? You've got
a robot for welding, a robot for
sorting, another one for vacuuming, and
each one has its own separate
specialized brain. But what if what if
we could give them all one single
unified brain? A brain that could learn
to do pretty much anything just by
understanding our world and our words.
Well, that's the VL revolution, and
we're going to break down how it is
changing absolutely everything. Imagine
saying that to a robot. Seriously, think
about it. Not some super specific
command like, "Pick up the green T-Rex
toy," but something that requires
abstract knowledge. Something it's never
ever heard before. And get this, this
isn't science fiction anymore. This is
the reality being built right now by a
new kind of AI. And it's giving robots a
power to understand our world in a way
we've only ever dreamed of. And this
this is the magic that makes it all
possible. The vision language action
model or VALA for short. It's such a
beautiful almost simple idea when you
break it down. It's one single model
that connects what a robot sees with its
cameras to what it understands from our
language to what it does with its body.
Vision, language, action. That trifecta
is what's finally making the dream of a
general purpose robot a reality. Okay,
so how in the world did we get to this
point? I mean, it didn't just happen
overnight. It all really kicked off with
two pioneering models that completely
shattered the old rules of robotics.
First up, you've got Google's RT2 back
in 2023. And honestly, this was the
field's Wright brothers moment. The
stroke of genius here was treating a
robot's physical actions, like moving an
arm to a specific spot, as if they were
just words in a sentence. I mean, how
clever is that? And that was
revolutionary because for the very first
time it allowed them to tap into the
massive knowledge of the entire internet
and connect it directly to physical
movement. Then in 2024 we got the Model
T moment with open VA. So if RT2 proved
that flight was possible, OpenVLA built
the affordable airplane that everyone
could use. As the first major open-
source model, it took this incredible
power and put it into the hands of
researchers and developers everywhere.
It was a total game changer. Now, this
is where it gets really really
interesting. Just look at the contrast.
Google's RT2 was this this behemoth,
right? 55 billion parameters. A true
proof of concept that needed massive
resources. But then look at OpenVLA. At
just 7 billion parameters, that's nearly
8 times smaller. It actually achieved a
16.5% higher success rate. This proved
that powerful robotics AI wasn't just
for the tech giants anymore. This
one-two punch is what lit the fuse. And
when I say it lit a fuse, I mean it led
to an absolute explosion of innovation.
A true Cambrian explosion for robotics.
After those pioneers laid all the
groundwork, the entire field just
erupted. The year 2025 is going to go
down in the history books for sure. Just
look at this timeline. For years,
progress was steady, but you know, kind
of slow. One model in 2022, the big one
RT2 in 2023, a handful in 2024, and then
boom, in 2025, the floodgates just burst
open with over 28 new models. I mean,
that is textbook exponential growth
right there on the screen. So, in total,
we went from just a couple of models to
over 35 in the span of 3 years. It's
just wild. And that created this
crowded, complex, and incredibly
exciting landscape. So the big question
becomes, how do we even begin to make
sense of it all? Well, we can actually
organize this whole explosion into three
key strategies, or what you could call
pathways to intelligence. Different
teams are tackling different core
challenges, pushing the boundaries in
their own unique ways. First up, we've
got the humanoid pathway. And this is
the grand challenge, right? giving a
robot with two arms and two legs that
fluid, coordinated, whole body control
it needs to operate in environments that
were built for us humans. This is
arguably the toughest nut to crack on
the hardware side of things. In this
table perfectly illustrates two totally
different approaches. On one hand, you
have figure AI's helix, which uses this
cool dual system brain. A slow,
thoughtful part for cognition and a
super fast 200Hz part for pure motor
control. But on the other hand, you have
Nvidia's GR0T using what's called a
frozen VLM plus adapter. So what does
that mean? Basically, they take a
massive pre-trained vision language
model and just lock it in place. That's
the frozen part. Then they add this tiny
trainable adapter to specialize it just
for robotics. It's an incredibly
efficient way to adapt a huge model.
Okay, our second path is all about
dexterity. It's one thing to move a big
arm around. It's another thing entirely
to master the delicate touch needed for
all those tasks that we do every day
without even thinking about it. So let's
look at a model like physical
intelligenc's pi0. This thing is a
master of manipulation. It uses a new
technique called flow matching which to
put it simply lets the model generate
these incredibly smooth and continuous
action commands instead of those jerky
discrete steps we're used to. And the
result? Well, it can fold laundry, bag
groceries, and assemble boxes. Tasks
that require a level of dexterity that
was pure science fiction just a couple
of years ago. Finally, we have the third
crucial path, efficiency. Because look,
all this incredible intelligence is
useless if it takes a data center to run
one robot. This pathway is all about
shrinking these powerful brains to fit
on affordable, accessible hardware that
can actually be deployed out in the real
world. And the progress here is just
staggering. Remember our pioneer RT2? 55
billion parameters. Now compare that to
a recent model called small VLA at just
450 million. That's over a 100 times
smaller. Yet it's powerful enough to run
realtime control on a single consumer
graphics card. The kind you could have
in your PC at home. This is what's going
to make widespread adoption a reality.
So what's the secret sauce driving this
incredible acceleration across all these
pathways? A huge part of the answer is
the open-source community which has
created this shared set of powerful free
building blocks that anyone can use.
Yeah, you can pretty much think of it
like a recipe. To build a modern VA, you
start with a powerful open-source vision
model like intern to act as the eyes.
You add a smart language model like
Llama 4 to be the cognitive core and
then you train it all on massive open
data sets of robot actions like the open
X embodiment data set. This open source
ecosystem is what's allowing the field
to move at such a breakneck pace. It's a
classic example of standing on the
shoulders of giants. All right, so let's
bring it all home. What does this all
mean for us? Where is this technology
actually taking us? This rapid leap
isn't just happening in a lab. It's
paving the way for a future where
intelligent robots are a part of our
daily lives. Now, of course, we're not
there yet. Let's be real. There are
major hurdles to overcome. We have to
ensure these robots are fundamentally
safe to be around. They need to be way
more robust to the chaos and
unpredictability of the real world and
the field still needs to find the best
most standardized ways to represent and
teach actions. The work is far far from
over. But the momentum is just
undeniable. This quote really captures
the feeling in the field right now. We
are on the cusp of creating truly
generalpurpose robots that can
understand our world, follow our
instructions, and work right alongside
us in our homes, our factories, and our
hospitals. And that leaves us with a
pretty profound question for the future,
doesn't it? We're moving towards a world
where robots can learn new skills, not
from complex code, but simply by
watching a video of a human doing a
task. And when that becomes commonplace,
what does it mean for the nature of
work, of skill, and of human endeavor
itself? That's something we're all going
to have to figure out together.