Sanctuary AI Phoenix — History & Background

Origins: Intelligence First

Sanctuary AI was founded in Vancouver in 2018 by Suzanne Gildert, Geordie Rose, and Olivia Norton. Gildert and Rose had previously worked together at D-Wave Systems, the quantum computing firm, and brought with them an unusual perspective on robotics. Rather than approaching the problem as a hardware challenge with AI bolted on — the dominant approach at the time — Sanctuary was founded on a specific thesis: that general-purpose physical intelligence is the bottleneck, and solving it requires building a new type of AI system from scratch.

The company named that system Carbon. Where most robotics AI in 2018 consisted of task-specific models, perception pipelines, and hand-coded control logic, Sanctuary wanted something more fundamental: a model that understood the world, could predict the consequences of physical actions, and could learn new tasks the way a person might — by watching and doing, not by being explicitly programmed.

The Carbon AI System

Carbon is best described as a large action model — an analogy Sanctuary draws to large language models, but with physical actions as outputs instead of text. It consists of two components: a world model that builds a representation of the robot's physical environment, and a task model that maps situations to actions. The system learns new tasks primarily through human teleoperation: an operator demonstrates a task remotely, controlling the robot's movements, and Carbon uses those demonstrations to build a generalised policy it can then execute autonomously.

This learning-from-demonstration approach distinguishes Phoenix from robots that require explicit programming or highly controlled environments. A task demonstrated once can sometimes be generalised to different object positions, lighting conditions, or slightly modified workflows without additional programming. In practice, Sanctuary claims Phoenix can perform over 110 distinct skills, spanning retail tasks like product handling and shelf-stocking to light manufacturing and healthcare support activities.

The approach has a practical limitation: it relies heavily on the quality and quantity of teleoperation data, and tasks that are difficult for humans to demonstrate at robot scale can be hard to teach. Sanctuary has been upfront about this, framing its commercial pilots partly as data-gathering exercises to keep improving Carbon.

From Prototype to Phoenix

The company spent 2018 to 2021 in relative stealth, building early prototypes and refining the Carbon architecture. The robot went through multiple internal iterations before Sanctuary was ready to show it publicly. When Phoenix was unveiled in 2022 — alongside a $58 million Series A — the company made clear that the hardware was designed around the AI, not the other way around. The hands, in particular, received significant engineering investment: by Generation 7 in 2024, each hand had 20 degrees of freedom, comparable in articulation to a human hand and surpassing most competitors at the time.

Generation 6 in 2023 was described by Sanctuary as the first commercially deployable iteration. Generations 1 through 5 were primarily internal research platforms. With Gen 6 came the Canadian Tire pilots — the first time Phoenix worked in a live commercial environment. Mark's (clothing and workwear) and Sport Chek (sporting goods), both part of the Canadian Tire family, hosted deployments where Phoenix performed retail tasks alongside human staff.

Commercial Reality and Competitive Position

Sanctuary's position in the humanoid market is distinctive. The company is smaller than Tesla, Boston Dynamics, or Figure AI in terms of funding and headcount, but it has something most competitors lack at the same stage: robots deployed in real paying-customer environments. The Canadian Tire and PriceSmart Foods pilots were commercial arrangements, not academic demos or controlled research projects.

That said, Sanctuary's deployment scale remains modest compared to the ambitions of American competitors. The company faces the challenge familiar to most humanoid startups: the hardware is expensive to produce, the AI needs enormous amounts of real-world data to improve, and enterprise buyers are cautious. Phoenix is not yet widely available for purchase — the company is focused on controlled expansion of its pilot programme rather than open sales.

Geographically, Sanctuary benefits from Canada's strong AI research ecosystem and government support for technology companies, though it competes primarily against well-funded US and Chinese players. The company has argued, with some credibility, that its emphasis on AI-first design gives it a different cost and scalability profile than hardware-first competitors who retrofit AI onto robots built to other specifications.