Sanctuary AI Built the World's First Robot That Thinks Before It Moves. 15 Units Is Just the Beginning.

Most humanoid robot companies start with a body and then try to give it a brain. Sanctuary AI started with a brain and then built a body around it.

That distinction sounds like marketing copy, but it describes a genuinely different engineering philosophy that makes Sanctuary one of the most intellectually interesting companies in the humanoid race. Operating out of Vancouver, British Columbia, the company has shipped 15 Phoenix humanoid robots. In a market where Unitree has shipped 5,500 and AgiBot has shipped 5,200, fifteen is barely a rounding error.

Sanctuary does not care.

The company’s founder, Geordie Rose, has spent his career building machines that think differently. He co-founded D-Wave, the first company to sell a commercial quantum computer. He co-founded Kindred AI, which built warehouse robots that learned by watching humans. Now he is building what he calls the world’s first human-like intelligence in a robot body, and he is willing to let every other company outship him while he gets the cognition right.

This is either a visionary bet that will eventually prove prescient, or a company that is falling further behind every quarter while competitors build insurmountable manufacturing leads. There is very little middle ground. And the evidence for both interpretations is compelling.

The quantum computing founder who pivoted to robots

You cannot understand Sanctuary AI without understanding Geordie Rose, and you cannot understand Geordie Rose without understanding D-Wave.

In 1999, Rose co-founded D-Wave Systems in Vancouver with the goal of building a practical quantum computer. At the time, most physicists believed quantum computing was decades away from any useful application. The academic consensus was that you needed thousands of perfectly coherent qubits to do anything interesting, and nobody could make even a few qubits work reliably.

Rose took a different approach. Instead of pursuing the theoretically optimal form of quantum computing (gate-model quantum computation), D-Wave built a quantum annealer, a less general but more immediately practical type of quantum computer. The machine could not do everything a theoretical quantum computer could do, but it could solve certain optimization problems faster than classical computers. D-Wave shipped its first commercial system to Lockheed Martin in 2011 and to Google and NASA in 2013.

The D-Wave experience taught Rose something that now shapes everything about Sanctuary AI: you do not need to solve the entire problem to ship something useful. You need to find the right abstraction layer, the right simplification, that lets you deliver real capability before the full theoretical framework is complete.

After D-Wave, Rose co-founded Kindred AI in 2014 with Suzanne Gildert, a physicist who had worked at D-Wave. Kindred built robotic picking systems for warehouse and logistics applications. The key innovation was a teleoperation-to-autonomy pipeline: human operators remotely controlled the robots initially, and the system learned from the human demonstrations over time, gradually taking over more tasks autonomously.

Kindred was acquired by Ocado Group in 2020. But the core insight from Kindred, that you could bootstrap robot intelligence from human demonstrations through teleoperation, followed Rose and Gildert to their next company.

Sanctuary AI was founded in 2018. The thesis was audacious: build a humanoid robot with human-like general intelligence. Not a robot that can do one task well. A robot that can understand what a task is, why it needs to be done, and how to adapt when things go wrong. Rose and Gildert believed this was possible because of what they had learned at Kindred about human-to-robot skill transfer. And they believed it was necessary because without genuine understanding, humanoid robots would forever be limited to narrow, pre-programmed applications.

Carbon: the AI system that comes first

The centerpiece of Sanctuary’s technology is Carbon, an AI control system that the company describes as a cognitive architecture for robots. Carbon is not a large language model bolted onto a robot. It is a purpose-built system designed from the ground up to give robots the ability to understand tasks at a conceptual level.

Most humanoid robot AI systems work roughly like this: the robot receives sensor data from cameras and other inputs, processes it through a neural network trained on specific tasks, and outputs motor commands. The system learns to map sensory inputs to motor outputs through massive amounts of training data. It can be very good at the specific tasks it was trained on, but it does not “understand” what it is doing in any meaningful sense.

Carbon works differently. The system attempts to model what Sanctuary calls the “task understanding layer,” an intermediate representation between raw sensory input and motor output that captures the conceptual structure of what the robot is trying to accomplish. When a Carbon-powered Phoenix robot is asked to fold a shirt, it does not just replay a memorized sequence of motor commands. It builds an internal model of what a shirt is, what “folded” means, where the fabric edges are, how to grasp and manipulate different materials, and how to recover when a fold goes wrong.

The distinction between Carbon’s approach and the reinforcement-learning or imitation-learning approaches used by most competitors is subtle but important. Most AI systems for robots learn to perform tasks. Carbon is designed to learn to understand tasks. The practical difference shows up when conditions change. A robot trained through standard imitation learning to pick up a red cup from a white table might fail when the cup is blue or the table is wooden. A robot with genuine task understanding should recognize that the color of the cup and the material of the table are irrelevant to the goal.

Whether Carbon actually achieves this level of task understanding is, honestly, still an open question. Sanctuary’s 15 deployed units have not generated enough public performance data to make a definitive judgment. The company’s claims are plausible based on the underlying research, but they have not been independently validated at the scale that would settle the debate.

The Phoenix robot

Phoenix is Sanctuary’s humanoid robot platform, now in its seventh generation. It stands 5 feet 7 inches tall, weighs approximately 155 pounds, and has a form factor that is deliberately human-scale. The robot is designed to work in environments built for people, fitting through standard doorways, reaching shelves at human height, and navigating workspaces with human-scale clearances.

The hands are where Sanctuary has invested disproportionate engineering effort. Phoenix’s hands feature multiple individually actuated fingers with force sensing and compliance control, capable of what the company calls “human-like dexterous manipulation.” In practical terms, this means Phoenix can handle soft objects, irregular shapes, and tasks that require fine motor control, capabilities that most competing humanoids lack or perform poorly.

This focus on hands is itself a philosophical statement. Sanctuary has argued publicly that hands are the most important and most difficult problem in humanoid robotics. Legs get more attention because bipedal walking is visually impressive and historically challenging. But in terms of commercial utility, the ability to manipulate objects with human-like dexterity is what determines whether a humanoid robot can actually replace human labor in most work environments.

Think about what human workers actually do in a retail store, in a factory, in a warehouse. They pick up objects of varying shapes and sizes. They open packages. They fold clothing. They assemble components. They type on keyboards. They handle tools. Nearly every commercially valuable task requires hands, and most of these tasks require the kind of fine manipulation that simple two-fingered grippers cannot perform.

Real deployments, small numbers

Sanctuary’s 15 shipped Phoenix units are deployed across a small number of pilot programs that reveal the company’s commercial strategy.

The most notable deployment is with Mark’s, a workwear and casual clothing retail chain owned by Canadian Tire Corporation. Phoenix robots have been tested in Mark’s retail locations performing tasks like folding clothing, organizing shelves, and managing inventory. Retail is a revealing choice for a first deployment. It requires general-purpose manipulation (handling many different types of objects), natural human-robot interaction (working alongside customers and staff), and adaptability (store layouts change frequently with seasons and promotions).

The more strategically significant partnership is with Magna International, one of the world’s largest automotive parts manufacturers. Magna produces components for virtually every major automaker and operates over 340 manufacturing operations across 28 countries. The partnership targets manufacturing tasks that require dexterous manipulation and cognitive flexibility, tasks that are currently performed by human workers because traditional industrial automation cannot handle the variability.

Sanctuary has also received financial support from the Canadian federal government, including approximately C$30 million in funding through various innovation programs. This government backing reflects Canada’s strategic interest in establishing a domestic AI and robotics industry, and it provides Sanctuary with non-dilutive capital that does not come with the same growth pressure as venture funding.

The total number of deployments is tiny. Fifteen units across a handful of locations does not generate the kind of operational data that Unitree’s 5,500 units or AgiBot’s 5,200 produce. This is the fundamental tension at the heart of Sanctuary’s strategy: the company believes it needs to get the intelligence right before scaling production, but getting the intelligence right requires real-world data that only comes from deploying at scale.

The contrarian argument

Every other humanoid company in the race has essentially converged on the same strategy: build the best robot body you can, ship as many units as possible, collect operational data, and improve the AI over time. The hardware comes first. The intelligence develops through deployment.

Unitree ships thousands of G1s at $16,000 each. The robots are not brilliant. They do not have sophisticated cognitive architectures. But each unit generates data, and that data feeds back into improvements. AgiBot follows the same playbook with its manufacturing-first approach. Tesla is leveraging FSD data and its own factory deployments. Figure AI is collecting data from BMW’s production floor.

Sanctuary inverts this. The company believes that shipping thousands of robots with limited intelligence is building on sand. You will get robots that can perform narrow tasks in controlled environments, but you will not get robots that can generalize, adapt, and truly replace human labor across diverse settings. The data you collect from narrow-task deployment teaches you how to do narrow tasks better. It does not teach you how to think.

Rose has been explicit about this framing. In interviews, he has compared the volume-first approach to trying to build a human brain by starting with reflexes and working upward. You can build very good reflexes. But reflexes are not intelligence, and no amount of reflex optimization will produce understanding. You need to build the cognitive architecture, the capacity for understanding, first, and then train it through experience.

The counterargument is equally compelling. You can theorize about cognitive architecture forever, but unless you deploy robots in the real world, you will never know whether your theories work. The messiness of real deployment, the unexpected failure modes, the edge cases that simulation never captures, is where true robustness comes from. Shipping 15 units generates 15 units worth of learning. Shipping 5,000 generates 5,000 units worth.

The teleoperation pipeline

One area where Sanctuary has genuine technical depth is in its teleoperation-to-autonomy pipeline, inherited and refined from the Kindred AI experience.

The pipeline works in stages. In Stage 1, human operators remotely control Phoenix robots in real time, performing the full range of tasks required in a deployment. The operators use VR-style interfaces with haptic feedback, feeling the forces that the robot’s hands experience and controlling the robot’s movements with their own body motions.

In Stage 2, the Carbon system analyzes the teleoperation data, identifying patterns, decision points, and strategies that the human operators used. This is not simple motion recording. The system attempts to extract the cognitive structure behind the actions: why did the operator choose to grasp the object at that angle? Why did they slow down at this point? What were they looking at when they made that decision?

In Stage 3, Carbon attempts to autonomously replicate the tasks that were demonstrated through teleoperation. The system starts with tasks where it has high confidence and gradually expands to more complex or less-certain tasks. When the system encounters a situation it cannot handle, it escalates to a human operator, and that intervention becomes additional training data.

This pipeline is theoretically elegant, but it has a practical limitation. Teleoperation is expensive. Every hour of robot operation in Stage 1 requires a human operator. The economics only work if Stage 3 autonomy develops fast enough to reduce the teleoperation burden before the costs become unsustainable. Sanctuary has not publicly disclosed what percentage of Phoenix operations are now autonomous versus teleoperated, which makes it difficult to assess how far along the pipeline has progressed.

The Canadian context

Sanctuary is the only major humanoid robot company headquartered in Canada. This matters for reasons that go beyond geography.

Canada has one of the strongest AI research ecosystems in the world. The country is home to Yoshua Bengio (University of Montreal), Geoffrey Hinton (formerly University of Toronto, now Google), and Richard Sutton (University of Alberta), three of the most influential figures in modern AI. The Vector Institute in Toronto, Mila in Montreal, and Amii in Edmonton form a national AI research network that produces world-class talent.

But Canada has historically struggled to convert its research strength into commercial success. Canadian AI startups frequently get acquired by American companies (Google acquired Hinton’s DNNresearch, Apple acquired multiple Canadian AI startups), and the country’s venture capital ecosystem is smaller and more conservative than Silicon Valley’s.

Sanctuary exists in this tension. It benefits from access to exceptional AI talent, government funding, and a supportive research environment. It is constrained by a venture ecosystem that cannot match the $1.85 billion that Figure AI raised or the effectively unlimited internal funding that Tesla can deploy.

The Canadian federal government’s investment in Sanctuary is a deliberate attempt to keep a strategically important AI robotics company in Canada. Whether C$30 million in government funding can substitute for the kind of venture capital available in the United States and China is an open question with historical precedent pointing toward no.

Why 15 units might matter more than 5,000

Here is the case for Sanctuary that its supporters make, and it deserves serious consideration rather than dismissal.

In the early days of personal computing, there were companies that shipped millions of units and companies that shipped thousands. The companies that shipped millions were running variants of CP/M and early DOS, operating systems that were functional but architecturally limited. Apple shipped far fewer Macintosh computers, but the Macintosh had a graphical user interface and a mouse, an architectural foundation that eventually became the standard for all personal computing.

The volume leaders had more data, more revenue, more market share, and more momentum. They lost anyway, because the architectural bet mattered more than the deployment numbers. The graphical user interface was not just a feature. It was a different way of thinking about what a computer should be.

Sanctuary is making a similar argument about humanoid robots. Carbon is not just a better AI. It is a different architecture for robot intelligence, one that prioritizes understanding over execution. If that architecture proves superior, then the thousands of robots deployed by competitors are training on the wrong paradigm. All the data in the world about how to perform narrow tasks does not help you build a system that can understand and generalize across all tasks.

The problem with this argument is timing. Apple’s graphical user interface advantage played out over a decade. In the humanoid market, the volume-first companies are not standing still. Tesla is integrating FSD-derived neural architectures. Figure AI is collaborating with OpenAI. AgiBot is deploying WorkGPT. The idea that only Sanctuary is working on intelligence while everyone else focuses exclusively on hardware is a straw man. Every humanoid company is investing in AI. The question is whether Sanctuary’s specific architectural approach to intelligence is sufficiently differentiated to overcome its massive deficit in scale, funding, and deployment.

The dark horse scenario

What would it look like if Sanctuary’s bet pays off?

The scenario goes something like this. Over the next two to three years, volume-first companies continue to ship thousands of units. These robots perform well on narrow, pre-defined tasks. They move totes in warehouses. They handle parts on assembly lines. They walk predetermined paths in factories. But they hit a ceiling. They cannot adapt to new tasks without extensive retraining. They cannot handle the unpredictable variability of retail environments, homes, or outdoor settings. They remain expensive, specialized industrial tools.

Meanwhile, Sanctuary reaches a threshold where Carbon’s cognitive architecture demonstrates genuine generalization. A Phoenix robot deployed in a Mark’s store proves it can handle not just clothing folding but inventory management, customer interaction, cleaning, and store reorganization, all from a single deployment with minimal task-specific training. The system transfers to a Magna factory and, with minimal adaptation, begins performing assembly tasks it was never explicitly trained for.

At that point, Sanctuary’s 15 units are worth more than a competitor’s 5,000, because each Sanctuary unit can do the work of multiple narrow-task robots. The economics flip. Customers stop buying five specialized robots for five tasks and start buying one general-purpose robot that can do all five.

In this scenario, Sanctuary becomes an acquisition target or a licensing platform. Magna, with its 340 operations in 28 countries, becomes the distribution channel. The Canadian government’s investment looks prescient. The company that shipped 15 units redefines the industry.

Is this scenario likely? That depends entirely on whether Carbon works as advertised. And that is something that cannot be determined from press releases or conference presentations. It can only be determined from years of real-world deployment data, which Sanctuary is only beginning to generate.

The honest assessment

Sanctuary AI occupies a position in the humanoid race that is simultaneously the most intellectually interesting and the most commercially precarious.

The intellectual interest comes from the genuine novelty of the approach. In an industry where most companies have converged on similar strategies, Sanctuary is running a fundamentally different experiment. If it works, it changes the game. If it does not, it was at least asking the right questions.

The commercial precariousness comes from the math. Fifteen units. One hundred to two hundred employees. C$140 million in total funding. No dedicated manufacturing facility. In the time it takes Sanctuary to ship its next 15 units, AgiBot will ship another few thousand. The gap is not closing. It is widening.

Sanctuary’s path forward requires one of two things to happen. Either Carbon demonstrates such compelling capabilities that a major partner (Magna, most likely) commits to large-scale deployment and provides the manufacturing resources Sanctuary lacks. Or a major technology company acquires Sanctuary for its AI technology and integrates Carbon into an existing hardware platform with established manufacturing scale.

Both paths are plausible. Neither is certain. And both require Sanctuary to survive long enough for Carbon to prove itself, which means the company needs to manage its burn rate carefully with limited funding and no clear timeline for when cognition-first will translate into revenue-generating scale.

What Sanctuary means for the race

Even if Sanctuary never ships more than a few hundred Phoenix robots, the company has already contributed something valuable to the humanoid industry: a clear articulation of what might be missing from the volume-first approach.

The question Sanctuary forces everyone to confront is this: what happens when narrow-task humanoids hit their ceiling? What happens when Amazon needs its warehouse robots to handle returns, not just move totes? What happens when BMW needs its factory humanoids to adapt to a new vehicle model without six months of retraining? What happens when the tasks run out and general capability becomes the minimum requirement?

Every company in the humanoid race will eventually need to answer that question. Sanctuary is the one company that put the answer first and the shipment numbers second. That makes it either the wisest company in the race or the most naive.

Fifteen units is not a market position. It is a research program with commercial ambitions. But research programs have a way of defining industries long after the companies that ran them have been absorbed into something larger.

D-Wave shipped the first commercial quantum computer. The company did not become the dominant force in quantum computing. But it proved the category was real, forced every major technology company to take quantum seriously, and created a market that grew far beyond what any one company could own. Rose has done this before. He may be doing it again.