Figure AI launches Helix: the AI system bringing humanoid robots into homes

Figure AI launches Helix: the AI system bringing humanoid robots into homes

🔎 Why 2026 is the year of the home robot

In February 2025, Brett Adcock announced the split between Figure AI and OpenAI. Two weeks later, he unveiled Helix, a proprietary AI model designed to operate a humanoid robot in a kitchen. The industry laughed. Twelve months later, nobody is laughing anymore.

Helix 02 just completed an autonomous kitchen task for 4 minutes without interruption, an unprecedented benchmark for humanoid robotics. Figure 03, the robot equipped with this "brain", is announced at ~$20,000 or $600/month to rent. Residential home trials are scheduled.

The question is no longer "will it happen" but "will the 2026 promises hold up against the reality of a living room". Let's analyze what Helix really changes, what Figure 03 can do today, and why the business model might be the real issue.

The essentials

• Helix is a VLA (Vision-Language-Action) model that unifies visual perception, language understanding, and motor control in a single neural network.
• Helix 02, its successor, adds a 3-layer architecture with a "System 2" dedicated to semantic reasoning, enabling 4 minutes of continuous autonomy in the kitchen.
• Figure 03 is priced at ~$20,000 for purchase or $600/month for rent, with at-home trials currently underway.
• Figure AI broke ties with OpenAI to develop Helix in-house, a massive technological gamble.
• Chinese competition (AGIBOT) is already shipping thousands of robots in mass production, putting pressure on Figure's timing.

Recommended tools

Helix: what this VLA model really is

Helix is not an LLM that you would just plug into a robot. It is a VLA (Vision-Language-Action) model designed from the ground up for robotic control. The difference is fundamental.

An LLM like GPT-5.5 reasons over text. It can generate instructions, plan steps, but it doesn't know how to drive actuators. Helix, on the other hand, integrates visual perception, language understanding, and motor control into a single neural network, according to the official description from Figure AI.

In practice, this means the robot doesn't go through an intermediary. It doesn't "translate" the voice command into text, then the text into an action plan, then the plan into motor commands. Everything is processed end-to-end in the same model. The result: reduced latency and more fluid coordination between what the robot sees, understands, and does.

Humanoid.guide describes Helix as a "world model" for humanoids. The robot doesn't just react to commands: it builds a representation of its environment and acts accordingly. This is what enables demonstrations of collaborative tasks between multiple robots, where each adapts its behavior to what the other is doing in real time.

Helix 02 and the 3-layer architecture

Helix 02, unveiled in 2026, takes the concept further with a three-layer architecture operating at different speeds, according to Techloy.

The fastest layer handles motor reflexes and balancing. The intermediate layer processes visual perception and motor planning. And the slowest layer, dubbed "System 2", handles semantic reasoning and language. This separation allows the robot to keep moving and reacting while it "thinks" about the next complex step.

It is this architecture that enabled the key demonstration: 4 minutes of autonomous kitchen task without interruption. Preparing a meal, manipulating objects, adapting to an unstructured environment. Geeky Gadgets points out that these 4 minutes represent a quantitative and qualitative leap compared to previous demonstrations in humanoid robotics, which usually lasted a few seconds.

Figure 03: the first consumer humanoid?

Figure 03 is the company's first robot explicitly designed for homes. The Figure 02 remained a research and industrial demonstration robot. Figure 03 changes the target.

According to Robozaps, the robot is priced at ~$20,000, a price that deliberately places it under the Tesla Optimus and in direct competition with the 1X NEO. Brett Adcock's idea is clear: make the humanoid accessible before Tesla locks down the market.

In terms of capabilities, The Decoder reports that Figure 03 is targeted at doing the dishes, floor cleaning, and general household chores, all via voice control. IBTimes UK confirms the ability to fold laundry and water plants.

Voice control is central. You say "clean up the kitchen" or "do the dishes", and Helix breaks down the task, identifies the objects, plans the movements, and executes. No smartphone, no complex interface. It's a bet on immediate accessibility.

Hardware specifications

Figure 03 retains humanoid bipedalism with hands equipped with fine grasping. Indian Express emphasizes that the robot is designed for mass production, not for the lab. This is an important detail: the difference between a prototype and a product is the ability to manufacture it in thousands of units.

The humanoid market is seeing two opposing approaches clash. On one side, affordable hardware coupled with imitation learning (Tesla). On the other, more expensive hardware with emergent reasoning (Figure). The debate over the right approach is far from settled.

Voice control and multi-robot coordination

Voice control is nothing new in itself. Alexa and Google Home have been doing it for years. But the difference with Helix is that the voice command doesn't trigger a pre-recorded script: it launches real-time reasoning.

When you tell Figure 03 "prepare a fruit salad bowl for me", the robot must identify the bowl, locate the fruits, assess their ripeness (visual perception), decide the order of manipulation (reasoning), and then execute the grasping and cutting gestures (action). All of this in a continuous flow, not in siloed steps.

AIBase reports that Brett Adcock presented Helix as the key to "improving the capabilities of humanoid robots in domestic tasks", with a particular emphasis on understanding natural instructions. No need to format your requests like SQL queries.

Multi-robot coordination is the other notable advancement. Geeky Gadgets documents demonstrations where two Figure robots collaborated on the same task in a kitchen. One prepared the ingredients, the other assembled the dish. This coordination requires each robot to model not only the static environment but also the intentions and movements of the other robot.

This is where the Skills system of other AI agents becomes interesting in comparison: in the software world, agents learn reusable "skills". Helix does something similar but in the physical world, learning action patterns that it can recombine according to the task.

The business model: $20,000 purchase or $600/month rental

This is perhaps the most critical point of the announcement. Figure offers two access models for Figure 03.

The ~$20,000 purchase (May 2026, check figure.ai) targets wealthy early adopters. IBTimes UK bluntly asks the question: is this a smart purchase or a toy for the rich? At $20,000, the robot must replace a part-time human cleaner for several years to be profitable. This is theoretically possible, but Figure 03's current capabilities probably don't justify it yet.

The $600/month rental changes the game. This model brings Figure closer to Robot-as-a-Service (RaaS), an approach that AGIBOT has already adopted on a massive scale. The advantage of renting: no upfront investment, software updates included, and most importantly, the risk is transferred to the provider. If the robot doesn't know how to wash your specific dishes, you just return it.

The problem is that $600/month is the price of a car. For a robot that, in 2026, will still be limited in its real-world capabilities. Interesting Engineering notes that the in-home trials of Figure 02 (alpha testing, late 2025) were accelerated thanks to Helix, but that the transition from the lab to a real living room remains the major challenge.

The "soon in homes" trap

Brett Adcock considers it possible to enter homes by 2026, but Indian Express reports that he himself acknowledges this is an "ambitious goal." This is a double-edged formulation.

The history of robotics is filled with "soon in homes" promises that never made it past the YouTube video stage. Jibo, Anki, Kuri: all announced the domestic revolution. All are dead. The difference with Figure is the level of capital (several billion dollars raised) and the quality of the underlying AI model. Helix is not a chatbot on wheels; it's a real control system.

But between a lab demo and 10,000 homes, there is an ocean of edge cases. Stairs, running children, the cat jumping on the robot, the sticky pan, the slippery floor. Each of these cases requires thousands of additional hours of training. The Skills system could help accelerate this learning, but the proof has not yet been established at a domestic scale.

The break with OpenAI: why and at what cost

The split between Figure AI and OpenAI, announced in early February 2025, surprised the industry. TechCrunch covered the event in detail: two weeks later, Helix was unveiled. The timing was not a coincidence.

The official reason: Figure wanted an AI model integrated into robotic control, not a generic LLM grafted onto a robot. OpenAI, with its models like GPT-5.5 or Claude Opus 4.7, excels in textual reasoning. But going from a token to a (force, position) pair for a robotic arm is a fundamentally different problem.

By developing Helix in-house, Figure controls the entire chain: from perception to action. This is a huge competitive advantage, but it's also a colossal risk. OpenAI spends billions on R&D for foundation models. Figure has to finance both the robotic hardware AND the AI software, with a fraction of the budget.

This divergence outlines a schism in the industry: those who believe that emerging reasoning in a VLA model will be enough (Figure), and those betting on cheap hardware with large-scale imitation learning (Tesla). Both approaches have merit. Neither has won yet.

The competition: AGIBOT crushes volumes, Genesis AI pushes full-stack

Figure is not alone in this race. And on one specific point — mass production — it is even lagging behind.

AGIBOT: the Chinese giant that doesn't make videos

AGIBOT is ranked #1 worldwide in humanoid robot shipments in 2025 by Omdia. Not in promises, in shipments. The company produced its 10,000th serial robot and launched a Robot-as-a-Service platform at MWC 2026 for €899/day in 17 countries.

Robonaissance details the strategy: AGIBOT deploys a facility called AIDEA for data collection and distributes units to universities. This is a "data-first" approach where every robot in service generates training data to improve the next one.

The €899/day RaaS seems expensive, but it's a B2B model. A factory pays €899 for a robot that works 24/7 without breaks or holidays. The return on investment is calculated in months, not years. Figure, with its $600/month in B2C, is targeting a radically different market.

Genesis AI and full-stack robotics

On the other end of the spectrum, Genesis AI dévoile GENE-26.5 et des mains robotiques humanoïdes : la robotique passe full-stack. Genesis's approach is to control the entire stack, from the AI model to the finger actuators. This is a similar positioning to Figure in principle, but with a different focus on manual dexterity.

The competitive landscape in 2026 is therefore shaping up into three blocks: the volume/Chinese approach (AGIBOT), the full-stack/reasoning approach (Figure, Genesis), and the minimal hardware/imitation approach (Tesla). None has dominated yet. All have serious arguments.

Comparison table: Figure 03 vs AGIBOT vs Tesla Optimus

Current limitations: what Figure 03 can't (yet) do

We must be honest: between the demonstration video and the robot in your kitchen, there is a gap that neither Helix nor any other model has yet bridged in 2026.

Figure 03's demos show a robot in a controlled environment, with objects arranged predictably. Your kitchen, with its stacked pots, trailing cables, and passing children, is a chaos that the robot does not yet reliably handle.

IBTimes UK cites critics who question "the real value at the current price and the early stage of development." This is harsh but not unjustified. At $20,000, one expects a mature product. Figure 03 is a pioneer product.

Energy autonomy is another rarely discussed topic. A humanoid doing the dishes, cleaning, and tidying up consumes a huge amount of energy. Figure does not communicate about Figure 03's battery life in real-world conditions. If the robot has to recharge every 45 minutes, its domestic utility drops drastically.

Finally, safety. A 60 kg robot moving around a home with children and pets must have an impeccable safety level. Regulatory standards for domestic robots are still embryonic. A single accident could hold back the entire industry for years.

❌ Common mistakes

Mistake 1: Confusing Helix with a classic LLM

Helix is not GPT-5.5 plugged into arms. It is a VLA model trained specifically to link perception, language, and motor action. Comparing it to an agentic LLM like Claude Opus 4.7 makes no sense: they do not solve the same problem. Helix produces motor commands, not text.

Mistake 2: Believing Figure 03 is ready for the general public

The "in-home trials" are alpha tests, not a commercial launch. Interesting Engineering makes it clear that this is a restricted program. Subscribing at $600/month today hoping for an autonomous cleaning robot means paying to be a beta tester.

Mistake 3: Ignoring AGIBOT on the pretext that it is Chinese

AGIBOT has shipped over 10,000 serial robots. Figure is still at the residential trial stage. Ignoring real production volumes in favor of American demo videos is an analysis error. Robotics is won in the factory, not on YouTube.

Mistake 4: Comparing B2B RaaS and B2C rental

€899/day at AGIBOT (B2B) and $600/month at Figure (B2C) are not comparable. The former is aimed at companies calculating an ROI on 24/7 operations. The latter targets individuals looking for home help. The business models have nothing to do with each other.

❓ Frequently Asked Questions

What is a VLA model?

A Vision-Language-Action model combines visual perception, language understanding, and motor control into a single neural network. Unlike an LLM that generates text, a VLA directly generates commands for the robot's actuators.

Does Helix work without an internet connection?

Figure has not officially communicated on this point. However, the architecture of Helix 02 with its three layers suggests that motor reflexes and perception can work locally, while semantic reasoning (System 2) might require cloud resources. To be confirmed.

Is Figure 03 really available for purchase?

No. The $20,000 announced by Robozaps is a "target price". Pre-orders are not yet open to the general public. Only in-home alpha testing programs are currently underway, according to Interesting Engineering.

Why did Figure leave OpenAI?

OpenAI develops generalist models (text, code, image). Figure needed a model integrated specifically for robotic control, where the output is not text but motor commands. The breakup allowed Figure to develop Helix as a dedicated VLA, according to TechCrunch.

How does Figure 03 compare to 1X NEO?

Both are targeting the domestic general public around $20,000. Figure 03 is betting on Helix and emerging reasoning. 1X NEO relies more on imitation learning. The 1X NEO is potentially closer to production, but Figure 03 has more advanced reasoning capabilities. The verdict is open.

✅ Conclusion

Helix is the best proof to date that an AI model can control a humanoid robot in a real domestic environment, with reasoning and not just imitation. But the 4 minutes of battery life in the kitchen remain a demonstration, not a product. Figure 03 at $20,000 or $600/month is a bold bet whose viability will depend on the ongoing in-home trials and the speed at which Figure can transition from the lab to the living room. To closely follow this humanoid robot race, reliable hosting like Hostinger remains the best starting point to build your own tech watch.
```