Internet BREAKS w⧸ The World’s Most Advanced AI Robot

Robotics and Automation
4 followers
AI Generated·May contain errors
Video Summary
AI GeneratedFigure introduces Helix 2.5, a breakthrough neural network enabling humanoid robots to perform household tasks in unfamiliar environments without site-specific training. By leveraging \"Index,\" a massive dataset of human experience, Figure has achieved zero-shot whole-body generalization, moving closer to a general physical intelligence for home and industrial use.
This is the most important project we've ever taken on at Figure.
To deploy robots at scale we need to deliver a robot to your home
and it has to work right out of the box. Prior to this we've been running Helix
autonomously but the data collected has been in each
environment. The breakthrough is we can generalize environments we've never seen
before and handle entirely new household
objects wherever they happen to be placed.
Today we're going to show you three tasks run by Helix 2.5
The first task is Figure 3 tidying living room.
My kids are constantly making a mess at home. I have toys scattered
everywhere. This is a home the robot's never been in before.
The robot's never seen any of these objects. We've made no changes to any of
the training. All we've done is basically book this
book this place, ship the robot here and now it's doing
real useful work. And the robot's constantly changing to
the dynamics of the environment. From here Figure 3 will go off and
basically clean the entire living room. All right we have another task to show
you now. Let's go. My dream is to never have to make my bed
again. In order to do that we have to be able
to deploy to any home. What's interesting here is Figure 3 has
never been in this room before. It's never seen this bed. It's never
seen this pillow and it has to be able to do autonomous work
fully end-to-end to make this bed. All right let me show you task three.
Okay this one's important. Doing laundry is the worst.
At my house we're doing laundry every single day.
This task is really difficult for robotics as it requires
really precise manipulation. With Helix 2.5
we're folding towels in a house the robot's never been in
with towels it's never seen. You just saw Helix do three tasks in this home
but we didn't just test here. We rented 30 homes across the Bay Area.
These are homes we never set foot in and we deployed Helix 2.5.
We saw success in every one of them. Now Corey
our head of AI will walk you through the results.
Helix 2.5 is a step change over all our previous systems.
It's the most advanced neural network we've ever built.
It's a single model that can tidy entire rooms, make beds,
and fold laundry in homes it's never been with items it's never been trained
on. These tasks were chosen because they
require solving nearly every classical robotics problem at once.
Perception, locomotion, manipulation, handling rigid and deformable
objects all in real time. This model doesn't just have to do this
in one home or two. We ran evaluations in 30 Bay Area
homes with zero data collected in any of them
and we saw Helix 2.5 succeed in all of them.
In robotics we call this zero shot whole body generalization
and it's the first result of its kind. What's really unique here is
you're seeing whole body humanoid behavior across entire rooms
in multiple homes with zero new training. All this takes an enormous amount of
prior experience so where does this come from?
Large language models really took off when people found a way to combine
scalable pre-training with massive amounts of data and compute.
Robotics doesn't have internet scale data so
a year ago we made a big bet. We launched Index,
a worldwide collection effort with a pretty simple idea.
Humanoid robots might actually be the only robot that has a loophole.
Their bodies move through the world the same way ours do so they might be able
to learn directly from human experience. Today over 90,000 people contribute every
week and 35 new minutes of human experience
gets uploaded to Index every second. All of this turns into training data for
Helix 2.5. The big question was does it actually
translate into robot behavior? So we ran a controlled experiment. We
took identical models. The only difference was one trained
without Index. It just didn't work. Only once we added
Index in to pre-training did we start to see
strong zero shot generalization over 30 homes.
Index was directly transferring behavior to the robot.
One of the most important lessons from LLMs were scaling laws.
How consistently next word prediction improved as you doubled data and compute.
In these experiments we found something quite similar. We found that next
robot action prediction was actually scaling similarly as you repeatedly
doubled Index. These are direct human to robot transfer
scaling laws, another first for humanoids.
Scaling was so smooth that we could actually predict our final run's
validation loss down to four decimal points before the training run ever even
started. That's early but it's exciting evidence
that robotic learning might be finally on a similar scaling path as LLMs.
So what does this mean? It doesn't mean robot learning is fully solved yet
but with Helix 2.5 we're starting to see the first signs of a more general
physical intelligence. We're committing to 3.5 billion dollars
of compute for Helix. Index training data is growing every day
and we now have a recipe that's turning all of this into robot behaviors.
This is figures path for getting a useful robot in every home.
It's time to scale up. I think humanoid robots will probably become the biggest
product category we've ever seen because we're talking
about a machine that could eventually do useful work for
almost every business and almost every family on planet earth
and the reason that matters becomes much clearer when you imagine somebody you
love needing help. Think about an elderly
parent who wants to stay in their own home.
They can still make their own decisions but changing the bedding,
carrying the shopping and keeping the place clean have become difficult.
A robot that reliably handles these jobs could help them stay independent for
longer. You're buying the ability to keep living
your own life. That also changes the life of their
children. A visit could become time spent together
instead of a growing list of jobs that needs doing and someone
trying to balance work, children and caring for a parent could get hours of
their week back. Multiply that need across millions of
families and you start understanding the potential demand.
But serving millions of different homes requires something that has been
extremely difficult for robots. They have to cope with the furniture, the
objects and the layout changes. That's why figures demonstration we
just watched is so interesting. Figure says it's
tested three household behaviors across 30
unfamiliar homes without training the robots in those
particular places. It's progressed towards a machine you
could put to work without rebuilding your home around it.
Now extend that idea to a hospital. A robot that can reliably collect
equipment, move supplies and restock cupboards
could take physical work off of the hands of busy staff.
Those jobs have to happen before anyone gets the benefit of a well-equipped ward.
If machines can handle more of them, trained people could spend more time
using the skills patients actually need. The economic benefit comes from making
those people's time go further and in elderly care that could mean a
carer spending longer with someone who needs attention
while a robot handles the washing or fetches supplies.
The World Health Organization projects that one in six people worldwide
will be 60 or older by 2030 so the need to support
independence is going to grow. There are also jobs
where the benefit goes beyond freeing up time.
Think about handling contaminated material or inspecting damaged equipment
somewhere unsafe for a person to enter. A suitably protected robot could reduce
the need to expose somebody to that danger.
It still has to survive the conditions and do the job properly
but a machine can be replaced in a way a person can't.
That same ability could become valuable when somebody needs rescuing.
After Fukushima, DARPA Robotics challenged testing machines
on tasks such as opening doors, climbing stairs and turning valves.
The aim was to help robots work in disaster environments too dangerous for people.
Future humanoids could help search damaged buildings
or take supplies to someone trapped inside.
Human-shaped bodies would be useful because of things like stairs,
handles and equipment which has been designed for the human form.
And that's how the opportunity expands from a household purchase
into something much bigger. The same basic robot design
could serve different customers as its software becomes more used to more tasks.
Manufacturers could spread development costs across a larger number of machines
potentially making each one more affordable.
Lower costs could then bring physical help within reach
of people and businesses that couldn't previously justify it.
A capable humanoid could earn its purchase price across several different jobs.
That's why its potential value could keep increasing after you buy it
provided new capabilities actually work on the hardware you own.
That's the sort of opportunity behind the enormous market forecasts.
In 2024 Goldman Sachs projected a $38 billion humanoid market by 2035
having substantially raised its earlier estimate.
Morgan Stanley's 2025 research looked further ahead
forecasting a market of roughly $5 trillion with around a billion humanoids in use.
As a revealing detail in Morgan Stanley's prediction
about 90% of those robots would do industrial and commercial work.
So its enormous forecast doesn't depend on everyone buying a robot for their home.
Businesses could drive much of the demand
provided the machines complete enough useful work to justify their cost.
That gives manufacturers a potential route to scale
while the more demanding household applications develop.
But we need to be honest about how much still has to improve.
Figure reported a 56% full task success rate in household evaluation
so plenty of tasks were still unfinished.
Something that struggles to finish chores
can't simply be promoted to look after a vulnerable person
and disaster response demand capabilities far beyond operating.
Reliability, safety and the cost of keeping those machines working
will determine which possibilities become products.
If manufacturers solve those problems we then face a challenge of our own.
The worker whose job disappeared doesn't automatically benefit
because their employer becomes more profitable.
We will need to decide how people share in the gains
and what support exists while jobs change.
That's why this conversation needs to reach people
who never watch a robot launch or read technology.
