Figure AI just revealed a massive $1 billion strategy to train humanoid robots for real-world work. Its Figure Index platform is collecting millions of human demonstrations to make Figure 03 smarter, faster, and more capable across factories, warehouses, and everyday environments.
At the same time, Figure 03 is showing progress in mobility, including autonomous ladder climbing, stair navigation, advanced vision, palm cameras, and full-body coordination powered by Helix AI. 🤖
Could Figure AI become a Tesla Bot killer and challenge Tesla Optimus Gen 3? The humanoid robot race is accelerating as US and Chinese robot companies push toward general-purpose AI robots.
Watch the full breakdown to see why Figure 03 could become one of the most important humanoid robots yet.
FIGURE has quietly revealed a massive plan to replace human workers with its FIGURE-04 robot,
and it comes with an aggressive strategy to invest over $1 billion into data and computing.
The company also paired this announcement with another demonstration showing FIGURE-03
climbing a vertical ladder to a raised platform and returning autonomously.
One update shows how aggressively FIGURE wants to build the intelligence needed to handle
everyday human jobs. The other shows how quickly the body is learning to move through physical
spaces built strictly for people. And the connection between those two developments
tells us much more about FIGURE's real direction than the hardware alone.
That is where FIGURE Index comes in. FIGURE introduced Index on August 25th
after running the platform quietly for around four months. Index pays people around the world
to record themselves completing real activities, creating physical world training data that can
be used to improve Helix. The platform has already reached more than 264,000 downloads
across 108 countries. More than 44,000 people contribute each week, and together they have
uploaded over 16 million videos. FIGURE has already paid around $15 million to the people
producing that data. The important part is the variety inside those recordings. Different people
can fold clothes differently, organize a kitchen differently, or handle the same object in completely
different ways. Index also includes tasks such as cleaning litter boxes, making beds, clearing
restaurant tables, serving customers, and stocking shelves. These are exactly the kinds of situations
a general-purpose humanoid could eventually encounter outside a controlled demonstration.
Instead of training Helix around only a small collection of carefully prepared tasks,
FIGURE can expose the AI to many environments, objects, and different ways humans solve the same
physical problem. The incoming footage goes through a pipeline that filters low-quality
recordings, checks for fraud, removes repeated material, balances the data set, and adds
descriptions to the accepted videos. This keeps the growing collection useful as thousands of
contributors continue adding new demonstrations. And FIGURE plans to make the Index much larger.
The company is preparing to expand the operation by roughly 100 times while investing more than
$1 billion into data and computing. Teaching a robot human tasks solves the intelligence side,
but the machine still needs to physically navigate where those jobs happen. The pace FIGURE is moving
at right now is wild, so definitely subscribe to the channel if you don't want to miss where this
goes next. That brings us to FIGURE 03's latest mobility test. On August 22, FIGURE founder Brett
Adcock shared footage of FIGURE 03 climbing a 15-foot industrial ladder to a raised platform
and then returning to the floor completely autonomously. Ladder climbing requires much
more coordination than normal walking because the arms and legs have to support the robot together
while every hand and foot moves toward the next rung. Each movement depends on the one before it.
A hand has to secure the next grip while a foot searches for another rung, and FIGURE 03 keeps
adjusting its balance as the body moves higher. Instead of letting the legs carry almost everything
like normal walking, the robot has to use its entire body to pull and support its own weight
through the climb. Coming back down creates an even harder problem. FIGURE 03 cannot clearly
watch its feet while they move toward rungs positioned behind and below its body. The robot
has to find those contact points while continuously transferring its weight between the hands and feet.
FIGURE 03 has cameras built directly into its palms for situations like this. When the arms
move above its head and the main cameras can no longer clearly see the hands, those palm cameras
provide another view of the ladder. Sensors in the fingertips can also detect extremely light
contact, which helps FIGURE 03 recognize when the hand has reached a rung before shifting more of
its body weight onto that grip. Then came the detail that made this demonstration much more
interesting. A paying customer actually needs FIGURE 03 to use this kind of ladder during work.
The robot has to climb to a mezzanine, complete its shift on the raised platform, and then make
the same journey back down. Would you trust FIGURE 03 working 15 feet above you in a real factory?
Tell me in the comments. That is important because many factories and warehouses were designed
entirely around human workers. Raised work areas can sit behind vertical ladders or steep stairways
rather than ramps. A robot limited to wheels would immediately lose access to those parts
of the building, while a humanoid with arms and legs can use the same route already built for
people. But the ladder was not the only mobility update. FIGURE also showed FIGURE 03 walking up
and down stairs, and this demonstration introduced an important change inside Helix. Its lowest
control layer can now combine the robot's sense of its own body with visual information from stereo
cameras. Previously, this layer depended mainly on proprioception. In simple terms, FIGURE 03
understood where its joints were positioned and how its body was balanced while moving.
Now the robot can also look at the terrain ahead and build a three-dimensional understanding of
where the ground, steps, and other surfaces are located. That gives FIGURE 03 more information
before placing each foot. When the robot approaches stairs or uneven ground, it can
read the shape of the surface and adjust where the next step should land. This creates more
stable movement across environments where the height or angle of the ground keeps changing.
FIGURE trained this behavior in simulation across many different terrains and conditions
before transferring it directly to the physical robot. That allows the AI to experience large
numbers of situations without repeatedly risking the hardware during training.
The ladder climb shows how quickly FIGURE 03 is becoming more capable inside spaces built for
people, while index shows where FIGURE wants to take that progress next. Better movement gets the
job, but the bigger challenge is giving it enough real-world experience to understand what needs to
happen once it gets there. FIGURE is now attacking both sides at once, improving the robot's body
while building the data needed to make that body genuinely useful in the real world.
And if you want to follow this race as these humanoids gain new abilities almost every week,
subscribe to NextGenHumanoids because this is moving much faster than most people realize.
Beijing hosted two major robot events in the same week,
and together they told the full story of where humanoid robotics stands right now.
The World Robot Conference filled four halls with more than 300 exhibitors and over 2,000
robotic exhibits. A short distance away, the World Humanoid Robot Games put more than 2,000 robots
into an arena and made them compete for real. UBTech gave the clearest picture of where realistic
humanoids are heading. The company brought its UWorld U1 robots to their first major public
showing, and visitors could walk straight up and start a conversation. UBTech displayed several
male and female versions, rather than one repeated face. Silicone skin and realistic
hair carried part of the effect, and precise eye tracking carried the rest. The lineup splits into
three levels. U1Light is a seated upper-body robot, starting near $17,600. U1Pro adds a full
walking body for flat indoor spaces. U1Ultra runs the company's Deeper Emotional AI and can
recognize more than 20 emotional states while remembering earlier conversations. The U1 is not
built to work. It is built to feel like another presence in the room. A head form proved how much
a face alone can pull a crowd. Its Elfchuan 2.0 stood on a stage and performed a live song while
reacting to the people watching. Her mouth followed the vocals almost syllable by syllable
as small movements passed across the eyes and cheeks. The newest version uses as many as 42
independent facial motors. The bigger change sits below the neck. A head form has added an
articulated upper body so Xuan can move her shoulders and arms while shifting posture.
That lets a gesture line up with eye contact, instead of forcing the head to do all the work.
The platform stayed fixed in place so the visible legs should not be read as walking ability.
Unitry pulled the largest crowds with a robot built for one purpose only.
Superman is a two-legged sprinter and nothing else. Unitry claims a top speed of 28.3 miles
per hour, which pushes past anything a human sprinter can hold. The robot also completed a
standing jump of roughly 6.5 feet. Both results come from the company rather than an independent
body. Superman has reportedly been in development for a little over three months. It has no hands
at all because the entire design goes into balance, landing, and raw movement.
Xiaomi ran the most complete task demonstration of the week by turning its CyberOne V2 into an
autonomous florist. The robot identified a flower, positioned itself at the workstation,
and picked the flower up. It added confetti, then turned to a visitor and handed over the
finished gift. Nobody was steering it through any of those steps. Think about how many separate
decisions you make wrapping a gift without noticing a single one. That is what Xiaomi
was actually showing. Those same hands have already spent four months inside Xiaomi's
electric vehicle factory, where the success rate on a self-tapping nut task climbed to 98%.
At Xsquare's booth, the main attraction had no face and no legs. Two robotic arms sorted
parcels without stopping while a live counter showed the total. An earlier unedited one-hour
run handled 1,816 parcels with better than 98% accuracy. Parcel induction is far harder than
moving identical boxes down a fixed path because soft bags slump against rigid packages,
and barcodes arrive facing every possible direction. The intelligence behind the arms
is Xsquare's Wall-B world action model, which finds a usable grip and frees one parcel before
rotating it until the barcode can be scanned. The rest of the floor filled in everything around the
humanoids. Galbit revealed ET1, its first fully bipedal robot, and put it on stage beside human
dancers every hour. Dax Robotics went somewhere else entirely with the QiGT-1000, a giant robot
horse rated to carry up to 2,205 pounds across snow and mud and rocky slopes. Everything on
that floor was a demonstration the companies could set up and control. The games removed
the safety net. The competition ran inside Beijing's National Speed Skating Oval,
with 16 countries entering and 51 events on the program. 30 were contests like racing,
fighting, and football. The other 21 placed robots inside simulated workplaces.
One rule shaped the entire event. A teleoperated run scored only half the normal points,
which pushed every team to let the robots work alone. Tiangong Ultra became the biggest star
on the track. The robot was developed by the Beijing Humanoid Robot Innovation Center and
ran the 100 meters in 9.39 seconds during a preliminary heat. Usain Bolt's human record
sits at 9.58 seconds, although humanoid races follow a different rulebook. Tiangong Ultra went
on to take the 400 meter title and later won the 1500 as well. Those longer races test far more
than top speed because the robot has to hold balance through turns and manage joint heat,
with nobody correcting a single mistake. Then came the part nobody planned. After one run,
Tiangong Ultra could not slow itself cleanly and pushed into a crash mat before stumbling
sideways past the line. Honor's Lightning Robot chased speed even harder. CCTV reported a peak
velocity of 14.5 meters per second, and Lightning finished its 100 meter heat in 9.47 seconds.
Then Lightning crossed the finish line and collapsed, and staff carried the robot away
on a stretcher. Driving a humanoid body forward at extreme speed is clearly becoming possible.
Bringing that same body back under control afterwards is still the harder problem,
so what matters more to you, raw speed or full control? Tiangong Omni won its 400 meter final
in the smaller size class, but the running posture drew nearly as much attention as the time. Omni
ran with both arms folded near the head instead of swinging them like a human athlete. Nobody
programmed that pose. It appeared during simulation training because the position eased load on the
shoulder joints and saved energy. Humanoid robots are built in our shape because our buildings were
built for us. That does not mean the best way to move a human-shaped body has to look human.
Galbut appeared at both events, and the arena version was the harder test by far. The company
put its robots on a tennis court against human players and says the robots completed more than
100 straight autonomous rallies. A sprinter always knows exactly where the finish line is.
A tennis robot has to track a ball that changes direction every second and move into position
before the shot arrives. One Galbut robot fell chasing a wide ball, then stood back up and kept
playing. Booster Robotics went after something quieter and far more valuable. During the opening
ceremony, the company sent 80 T2 humanoids onto the track together in one long formation. A single
robot completing a hard move proves that one unit can do it. 80 robots moving in step proves the
same hardware can be built and calibrated across an entire fleet. The quieter events mattered most.
Robots were sent into simulated workplaces to connect cables, charge electric vehicles,
and tighten screws. Tightening a screw sounds trivial until a robot has to do it alone. It has
to find the screw and line the hand up with the right amount of force. Your hands solve all of
that in a second without any conscious thought. Robots crashed and fell all week, and that is
exactly why open competition is useful. A company can edit a demonstration before publishing it.
An arena full of cameras leaves far less room to hide. Chinese companies filled almost every track,
court, and booth across both events. Major American humanoid companies were largely absent,
and that gap is a choice rather than a scoreboard. Chinese firms are putting robots into public
competition where success and failure are equally visible, while American firms push
harder on factory pilots and paid deployments behind closed doors. Put both events side by side
and the picture gets clear. The conference showed what these robots can present, from a singing face
to arms sorting thousands of parcels. The games showed what happens when nobody is allowed to
reset the demo. Humanoid robots are now fast, expressive, and genuinely capable at real work.
The problem that remains is reliability, and it appears every time a robot crosses a line,
it cannot stop behind. If you want to keep up with how quickly this is changing,