Figure 03

Open on YouTube

External video · Figure AI on YouTube

Figure 03 is the third-generation humanoid robot developed by Figure AI, designed to combine general-purpose physical intelligence with a humanoid form factor suitable for both homes and commercial environments.

Unlike humanoid robots that are primarily presented as industrial automation machines, Figure 03 is being developed around a broader ambition:

a general-purpose robot capable of learning and performing everyday human tasks.

At 5 feet 8 inches tall, approximately 61 kilograms, and fully electric, Figure 03 is designed to operate in environments built for people rather than specially engineered robotic spaces.

Its current development is closely tied to Helix, Figure's proprietary vision-language-action AI system. The combination of Figure 03's new sensor suite, dexterous hands, tactile sensing, onboard computing, and Helix is intended to allow the robot to understand its surroundings, interpret natural-language instructions, and execute physical tasks.

Figure 03 also represents an important shift in the company's strategy.

Earlier Figure robots demonstrated the potential of humanoids in industrial environments.

Figure 03 was engineered from the beginning with three larger goals:

AI capability.

Operation in the home.

High-volume manufacturing.

That combination makes Figure 03 one of the most ambitious attempts to turn humanoid robotics into a scalable general-purpose product.

Figure 03 at a Glance

Figure 03 is a full-size electric humanoid robot designed for general-purpose physical work.

The basic idea is straightforward:

Humans live and work in environments designed around the human body.

We build:

• Kitchens
• Stairs
• Doors
• Cabinets
• Shelves
• Washing machines
• Dishwashers
• Tables
• Tools
• Workstations

A humanoid robot can potentially use the same physical infrastructure without requiring an entirely new environment.

Figure 03 takes that idea further by combining the humanoid body with an AI system intended to understand natural-language instructions and adapt to unfamiliar situations.

Instead of programming every action manually, Figure's long-term objective is to create a robot that can be told what to do and then figure out how to do it.

For example:

"Put the dishes away."

"Move the laundry to the dryer."

"Clean the kitchen."

The robot's intelligence must translate these high-level instructions into a sequence of physical actions.

That is the problem Figure 03 is designed to address.

The most important feature of Figure 03 is not its height or maximum speed.

It is the relationship between hardware and AI.

Humanoid robotics has traditionally been constrained by a fundamental problem:

A robot can be mechanically impressive while remaining difficult to program.

Every new task may require:

• New motion planning
• New perception logic
• New object models
• New manipulation strategies
• New control policies

That approach becomes extremely difficult to scale.

The home makes the problem even harder.

A factory can be carefully organized.

A home cannot.

A kitchen may contain hundreds of objects.

Laundry may be folded, crumpled, wet, or scattered.

Children may leave toys on the floor.

Furniture may move.

People may suddenly walk into the robot's path.

Figure's answer is to make the robot increasingly dependent on learned AI rather than manually engineered task-specific behavior.

This is where Helix becomes central to Figure 03.

Figure 03 was designed specifically around Figure's Helix Vision-Language-Action system.

Helix connects three traditionally separate capabilities:

Perception

The robot understands what it sees.

Language

The robot understands what a person is asking it to do.

Action

The robot converts that understanding into physical movement.

Figure introduced Helix as a generalist VLA model designed to connect visual and linguistic understanding with learned robot control.

The significance is substantial.

A traditional robot might be programmed with:

If object A is detected at position X, move arm to position Y.

A generalist AI robot aims for something closer to:

"Pick up the red cup."

The robot then determines:

• Which object is the cup
• Where it is
• How to reach it
• How to grasp it
• How much force to use
• Where to place it

This is a much more scalable approach to robotics.

Figure's AI system has continued to evolve alongside Figure 03.

In 2026, Figure introduced Helix 02, extending the architecture toward full-body autonomy. The system takes input from head cameras, palm cameras, fingertip tactile sensors, and the robot's proprioceptive state, while producing joint-level control across the body.

This is important because humanoid manipulation cannot be separated cleanly into "arms" and "legs."

Consider picking up an object from a low cabinet.

The robot may need to:

1. Walk toward the cabinet.
2. Stop at the correct distance.
3. Bend its knees.
4. Adjust its torso.
5. Extend its arm.
6. Reach into the cabinet.
7. Locate the object.
8. Grasp it.
9. Maintain balance.
10. Pull the object out.
11. Stand upright.
12. Carry it somewhere else.

The entire body becomes part of the manipulation system.

Figure's move toward full-body control therefore represents a significant evolution beyond simple arm manipulation.

Figure 03 includes a substantially redesigned sensory system.

According to Figure, its new camera architecture provides:

• Twice the frame rate
• One-quarter the latency
• 60% wider field of view per camera
• Expanded depth of field

These changes are intended to provide Helix with a denser and more stable stream of visual information for navigation and manipulation.

For a humanoid operating in a home, this matters enormously.

A robot may need to recognize objects while:

• Walking
• Turning
• Reaching
• Crouching
• Carrying something
• Working in a confined space

Higher-frequency and lower-latency perception can help reduce the gap between seeing an event and reacting to it.

One of Figure 03's most interesting hardware changes is the addition of cameras directly inside the palms.

This solves a practical problem.

When a robot reaches into:

• A cabinet
• A refrigerator
• A drawer
• A box
• A crowded shelf

its head-mounted cameras can become partially or completely blocked.

Palm cameras provide another source of visual information from much closer to the object being manipulated.

This allows Helix to maintain visual awareness during close-range manipulation.

It also creates a useful redundancy:

Head cameras → understand the environment

Palm cameras → understand the immediate manipulation area

The two levels of perception can work together.

Figure 03 also introduces a redesigned hand system with softer, more compliant fingertips and internally developed tactile sensors.

Each fingertip sensor can reportedly detect forces as small as 3 grams. Figure compares this sensitivity to detecting the weight of a paperclip resting on a fingertip.

This level of sensitivity matters because manipulation is not simply about seeing an object.

The robot must also feel contact.

Consider picking up:

• An egg
• A glass
• A thin plastic bag
• A paperclip
• A piece of clothing

Too little force can cause the object to slip.

Too much force can damage it.

Tactile feedback provides another layer of information that vision alone cannot provide.

Vision can tell the robot:

"I see the cup."

Tactile sensing can tell the robot:

"I am holding the cup."

That distinction becomes important during real-world manipulation.

An object may be:

• Slippery
• Flexible
• Partially hidden
• Irregularly shaped
• Moving
• Deformable

Tactile feedback can help the robot continuously adjust its grip.

This is one of the technologies that could eventually make humanoids significantly more capable in homes.

Figure 03 has a listed payload of 20 kilograms.

This gives the robot sufficient capacity for many common household and commercial objects.

The figure should not be interpreted as meaning that Figure 03 can safely carry 20 kg in every posture or situation.

Real-world payload capability depends on:

• Object shape
• Center of gravity
• Distance from the body
• Walking speed
• Grip
• Joint configuration
• Environment

Nevertheless, the specification places Figure 03 well beyond the category of a purely lightweight demonstration robot.

Figure currently lists a maximum speed of approximately 1.2 m/s.

This is not a particularly important specification by itself.

A general-purpose humanoid does not need to run extremely fast inside a home.

Instead, useful mobility depends on:

• Stability
• Precision
• Navigation
• Turning
• Obstacle avoidance
• Recovery
• Coordination with manipulation

The important question is not:

"How fast can it run?"

It is:

"Can it move naturally through a house while carrying out another task?"

That is a much harder robotics problem.

Figure 03 is listed with a runtime of approximately 5 hours.

The underlying F.03 battery uses a 2.3 kWh energy pack.

Figure says the battery was designed to provide:

• Approximately 5 hours of runtime
• 2 kW fast charging
• Active cooling
• Improved energy density
• Improved safety
• Lower cost compared with the previous generation

Five hours is particularly significant for a household robot because the robot does not necessarily need to operate continuously.

It can perform tasks during the day and recharge when necessary.

Figure 03 supports wireless inductive charging.

Charging coils are integrated into the robot's feet, allowing the robot to step onto a charging stand and recharge without plugging in a physical cable. Figure lists charging at up to 2 kW.

This is a deceptively important feature.

Imagine a household robot that requires someone to manually connect a charging cable every few hours.

That creates a significant usability problem.

Wireless charging allows a more autonomous cycle:

Work → Return to dock → Charge → Resume work

The robot does not need to depend on a human for routine charging.

The home is arguably the most difficult environment for a general-purpose robot.

Factories are designed around repeatability.

Homes are designed around people.

A household environment contains:

• Unpredictable object placement
• Soft materials
• Small objects
• Fragile objects
• Narrow spaces
• Stairs
• Pets
• Children
• Moving people
• Changing layouts

Figure 03 was therefore redesigned specifically with home operation in mind.

The robot uses softer external materials and strategically placed multi-density foam to reduce hazards around contact and pinch points.

Figure 03 is covered with soft textiles rather than exposed hard mechanical components.

This makes sense for a robot expected to work around people.

In a factory, exposed mechanical structures may be acceptable.

In a home, a robot constantly moving its arms and body near people needs a different design philosophy.

The exterior therefore becomes part of the safety system.

Figure also says the soft goods are:

• Washable
• Removable
• Replaceable without tools

This makes the robot more practical for everyday environments.

Figure states that Figure 03 has approximately 9% less mass and significantly less volume than Figure 02.

This is important for household operation.

A smaller, lighter robot can potentially:

• Move through tighter spaces
• Navigate around furniture
• Reduce collision energy
• Make handling easier
• Improve overall usability

The objective is not simply to make the robot smaller.

It is to make the humanoid body more compatible with real environments.

Figure's current product positioning specifically highlights autonomous household tasks including:

• Laundry
• Cleaning
• Dishwashing

These tasks are interesting because they are not standardized industrial operations.

Laundry is highly variable.

Clothing can have:

• Different shapes
• Different textures
• Different weights
• Different levels of deformation

Dishes introduce another challenge.

They can be:

• Fragile
• Slippery
• Wet
• Different sizes

A robot capable of handling these objects reliably needs a combination of perception, tactile feedback, dexterity, balance, and planning.

Figure 03 is designed around a simple interaction model:

Talk. Ask. Delegate.

The objective is to allow people to communicate with the robot using ordinary language rather than a dedicated programming interface.

This could dramatically change how people interact with robots.

Instead of:

Select task → Configure robot → Start program

the interface becomes:

"Can you clean the kitchen?"

The AI then needs to determine what "clean the kitchen" actually means in the current environment.

This is one of the major ambitions behind Helix.

A powerful general-purpose robot needs to separate what the human wants from how the robot accomplishes it.

A person might say:

"Put these groceries away."

The human does not need to specify:

• Which hand to use
• Where to stand
• How to grasp each package
• How to open the refrigerator
• How to move around the kitchen

The robot needs to determine those details itself.

This distinction is fundamental to general-purpose robotics.

Humans communicate goals.

Robots need to figure out actions.

Figure 03 is being developed around that model.

Figure 03 supports 10 Gbps mmWave data offload.

Figure says this allows the robot fleet to upload very large amounts of data for continuous learning and improvement.

This is important because modern robot AI depends heavily on data.

Every robot interaction potentially generates information about:

• Objects
• Human behavior
• Successful grasps
• Failed grasps
• Navigation
• Motion
• Contact
• Environmental changes

The more robots deployed, the more data can potentially be collected.

That creates a feedback loop:

More robots → More data → Better models → Better robots → More deployment

This may ultimately be one of Figure's most important competitive advantages.

One of Figure 03's biggest differences from earlier humanoid robots is that it was designed for high-volume manufacturing from the beginning.

This is a critical distinction.

Many humanoid robots are effectively engineering prototypes.

They may use:

• CNC-machined components
• Low-volume parts
• Custom electronics
• Hand assembly
• Expensive manufacturing processes

Those approaches are useful for development but difficult to scale.

Figure redesigned Figure 03 with manufacturing economics in mind.

Figure says Figure 03 relies heavily on manufacturing processes such as:

• Die casting
• Injection molding
• Stamping
• Metal injection molding

rather than relying primarily on CNC machining.

This is a major industrial-design decision.

CNC machining is flexible but can become expensive and slow at very high volumes.

Tooling-based processes require greater upfront investment but can dramatically reduce the marginal cost of each unit once production reaches scale.

This suggests Figure is not thinking about Figure 03 as a limited-run research machine.

It is thinking about it as a mass-produced consumer and commercial product.

Figure created BotQ, its dedicated humanoid manufacturing facility, to support high-volume production.

Figure says the first-generation BotQ production line is designed to produce up to 12,000 humanoid robots per year, with a goal of producing 100,000 robots over four years.

The significance of BotQ goes beyond manufacturing capacity.

Humanoid robots require a supply chain that is still relatively immature.

Figure has therefore moved toward vertical integration of important components including:

• Actuators
• Batteries
• Sensors
• Structures
• Electronics
• Hands
• Final assembly

This gives Figure greater control over:

• Cost
• Quality
• Manufacturing iteration
• Supply
• Product architecture

Figure has also developed the F.03 battery internally.

Figure says the battery represents the third generation of its battery architecture, with a 94% increase in energy density from the original generation and a 78% reduction in cost compared with F.02.

The battery is integrated directly into the torso.

This reduces the need for bulky external battery hardware and allows the battery structure to contribute to the robot's overall mechanical architecture.

The battery also incorporates multiple levels of protection at the:

• Battery-management-system level
• Cell level
• Interconnect level
• Pack level

Figure says the battery has achieved UN38.3 certification.

The manufacturing strategy is one of the most important aspects of Figure 03.

Figure is not simply trying to make a better humanoid.

It is trying to solve a larger problem:

How do you manufacture enough humanoid robots for general-purpose AI to become economically useful?

A robot intelligence platform needs physical distribution.

One robot produces limited data.

Ten robots produce more.

One thousand robots produce dramatically more.

At large scale, improvements in AI can potentially be deployed across the entire fleet.

This is why manufacturing capacity is closely connected to Figure's AI strategy.

Although Figure 03 is now strongly associated with the home, its capabilities are also relevant to commercial environments.

Figure says the same improvements developed for household environments can benefit commercial applications.

These include:

• Faster actuators
• Better perception
• Improved manipulation
• Tactile sensing
• Palm cameras
• Wireless charging
• High-speed data transfer

This creates an interesting strategy.

Instead of building one robot for factories and another for homes, Figure is attempting to build a general-purpose platform capable of operating in both.

A factory typically controls:

• Lighting
• Layout
• Object placement
• Workflow
• Human access
• Robot routes

A home controls none of these perfectly.

The environment changes constantly.

This makes the home a much harder test of general-purpose physical intelligence.

If a robot can reliably perform useful tasks in a cluttered household, its underlying capabilities may transfer to many commercial environments.

That is why Figure's home strategy is more ambitious than it may initially appear.

Figure 03 is aimed at a much broader audience than a conventional industrial robot.

Potential applications include:

Home Assistance

• Laundry
• Cleaning
• Dish handling
• Household organization
• Object retrieval

Commercial Work

• Logistics
• Material handling
• Packaging
• Manufacturing
• Warehouse operations

Research

• Embodied AI
• Vision-language-action models
• Robot learning
• Dexterous manipulation
• Human-robot interaction

Future Consumer Robotics

The long-term vision is a robot that ordinary people can interact with using natural language rather than specialized robotics software.

The household robot has been a long-standing dream in robotics.

But previous generations struggled with one fundamental problem:

Homes are too complicated.

A robot may work perfectly in a laboratory and fail when faced with:

• A tangled shirt
• A half-open drawer
• A transparent cup
• A toy on the floor
• A crowded countertop
• A new brand of packaging

This is why Figure's AI strategy is so important.

A robot cannot be manually programmed for every object and every household.

It needs to generalize.

It needs to understand new situations.

It needs to learn from experience.

And it needs to translate natural-language goals into physical actions.

Figure 03 is being designed around exactly this problem.

Figure's strategy can be understood as a loop:

Robot → Data → AI → Better Robot → More Robots → More Data

This is fundamentally different from traditional robotics.

Traditional robotics often scales through engineering:

More engineers → More programmed tasks

Figure is attempting to scale through AI:

More robots → More data → Better models → More general capabilities

This is why the company's manufacturing strategy and AI strategy are deeply connected.

The humanoid robotics industry is increasingly divided into several approaches.

Some companies emphasize:

Dynamic movement

Others emphasize:

Industrial automation

Others emphasize:

Low-cost hardware

Others emphasize:

Open robotics platforms

Figure is taking a different route:

General-purpose physical intelligence.

The objective is to make the robot increasingly capable of understanding the world and acting within it.

That makes Figure 03 less like a conventional industrial machine and more like a physical AI platform.