Sanctuary AI builds Physical AI systems that combine vision, touch, and learned decision-making to automate industrial tasks involving objects that bend, shift, or vary across cycles. The core technology stack, originally developed as Carbon for the Phoenix humanoid, also runs on commercially available industrial robot arms from manufacturers including FANUC and Universal Robots.

For a typical deployment, Sanctuary and the customer identify a repetitive task that conventional fixed automation cannot perform reliably, such as inserting flexible wired connectors into moving targets on a conveyor. Sanctuary configures a robot arm with an off-the-shelf or custom end effector, collects training data through human demonstrations and simulation, and trains an autonomous task policy using reinforcement learning and real-world validation. Cameras locate objects, tactile and force sensors detect contact conditions, and AI models adjust the robot's motion rather than replaying a fixed path.

Sanctuary's proprietary hydraulic robotic hand is a five-finger anthropomorphic gripper with integrated tactile sensing and 17 actively controlled degrees of freedom. It can rotate and reorient objects within the hand without setting them down. Policies trained through simulation in NVIDIA Isaac Lab transfer directly to physical hardware with zero additional real-world retraining. The hand remains in advanced development and is not yet generally available for production use.

Phoenix, now in its eighth generation, uses a wheeled base instead of bipedal legs after customer feedback indicated that contemporary legs were too fragile for the strong, precise torso required in factory work. It functions primarily as a research, data-capture, and teleoperation platform that supplies training data to the Physical AI stack, rather than as a mass-produced commercial product.

Sanctuary plans to add autonomous mobile robot bases for tasks requiring movement between workstations, with pilots expected soon. Its product range extends from fixed industrial arms with simple grippers to custom end effectors, dexterous hydraulic hands, mobile manipulation, and eventually full humanoid configurations, all controlled by the same Physical AI intelligence layer.

Sanctuary AI sells enterprises a full-stack automation system combining AI task-policy development, systems integration, robot hardware configuration, and deployment engineering. Its near-term commercial unit is an intelligent automated work cell rather than a packaged humanoid robot, allowing the company to generate revenue using third-party industrial arms without waiting for Phoenix to reach mass production.

The go-to-market motion follows a consultative land-and-expand model. Sanctuary starts with a narrowly scoped, high-value automation problem, tests performance against the customer's production benchmarks, and then targets adjacent tasks in the same facility and across other sites. For example, a validated task policy for connector insertion could be replicated at dozens of comparable plants with less engineering than the first installation required.

Revenue today likely combines non-recurring engineering fees, hardware and end-effector configuration charges, and deployment milestones. Recurring revenue could eventually include per-cell software licenses, annual support and maintenance, model retraining packages, and fleet-management services, though Sanctuary has not disclosed standard pricing. Its cost structure includes high fixed R&D spending on AI research, hydraulic-hand development, simulation infrastructure, and specialized talent, plus variable deployment costs for on-site engineering, custom tooling, data collection, and commissioning.

Deploying Physical AI on existing industrial robots could reduce capital requirements. Rather than simultaneously addressing locomotion, balance, battery life, actuators, manufacturing, and service infrastructure, Sanctuary can focus spending on perception, AI, task policies, and dexterous end effectors while using the mature arms, safety systems, and service networks of established robot OEMs. Margins could improve if Sanctuary reuses task policies across customers, standardizes work-cell designs, reduces customer-specific data requirements, and converts deployments into recurring software contracts. In the near term, however, the model resembles a services-heavy integration business.

Sanctuary AI competes in two overlapping markets: full-stack humanoid robotics and hardware-agnostic Physical AI software. Its June 2026 pivot to deploying Carbon on existing industrial arms puts it against both humanoid OEMs and embodied-AI platforms.

Figure AI is the most heavily funded Western competitor, with more than $1.75B raised at a $39B valuation, a BotQ factory designed for 12,000 robots annually, and a production deployment at BMW where Figure 02 logged over 1,250 hours and loaded more than 90,000 parts. Figure is also building vision-language-action models through Helix and securing access to as many as 100,000 NVIDIA GPUs, creating a data and compute advantage over Sanctuary.

Tesla Optimus poses a structural threat because Tesla can combine automotive-scale manufacturing, captive factory deployment, battery and actuator supply chains, and autonomous-driving perception talent. Optimus does not need to win external customers initially. It can accumulate operating data and reduce unit costs inside Tesla's plants.

Agility Robotics has the most extensive publicly documented commercial record among humanoid peers, with Digit 4 accumulating over 65,000 operating hours and more than $300M in multiyear orders. Its robot-as-a-service agreement with GXO and status as NVIDIA's initial partner for the Halos safety stack give it procurement and certification advantages that Sanctuary has not matched.

Apptronik closed more than $935M in funding and combines modular Apollo hardware with Google DeepMind's Gemini Robotics for AI, Jabil for contract manufacturing, and Mercedes-Benz for deployment. Its ecosystem strategy resembles Sanctuary's hardware-agnostic direction, but Apptronik has substantially more capital and a direct relationship with Google DeepMind.

Google DeepMind's Gemini Robotics is explicitly multi-embodiment and can control conventional dual-arm systems and Apptronik humanoids with different hand configurations. If robot OEMs can license a broadly capable Gemini model, Sanctuary must show that Carbon delivers better task success, cycle time, and tactile control on contact-rich work, rather than competing only on general reasoning.

Physical Intelligence and Skild AI, the latter valued at $14B on $1.8B raised, are developing hardware-agnostic foundation models for robot control that could commoditize Carbon's software layer. Chinese manufacturers including Unitree, AgiBot, Fourier, and Galbot are setting the hardware cost floor, with Unitree listing the G1 from approximately $13,500 and Chinese vendors reportedly accounting for more than two-thirds of the roughly 15,000 humanoids shipped globally in 2025. U.S. protectionism may limit Chinese penetration in sensitive industrial settings, but it does not eliminate downward pressure on hardware pricing.

Sanctuary AI's primary TAM expansion is the shift from selling complete Phoenix humanoids to deploying Physical AI on the five million industrial robots operating in factories globally. Proprietary dexterous hands and mobile manipulation add further product layers.

By running Carbon on installed FANUC, Universal Robots, and other third-party arms, Sanctuary can address the existing base of industrial automation without requiring customers to adopt a new robot platform. This converts incumbent robot fleets from substitutes into potential distribution infrastructure.

Compatibility with established hardware also creates an indirect channel through systems integrators and robot distributors. Sanctuary can access their installed bases and service organizations without building its own global field-support network.

Sanctuary's hydraulic hand combines tactile sensing, high force density, and fine control. Sold as a standalone end-effector-plus-intelligence package, it could serve customers that need capabilities beyond conventional two-finger grippers, including wire harnesses, connectors, flexible materials, small-part assembly, and tool use.

The Zeon materials partnership targets a barrier to commercializing dexterous hands at scale: maintaining tactile sensitivity through millions of high-cycle industrial movements. If Sanctuary productizes the hand as a platform, it could generate revenue and training data when customers choose another robot body.

Sanctuary's initial automotive focus could extend into electronics and semiconductor manufacturing, pharmaceutical logistics, and warehousing. Labor shortages, ergonomic constraints, and high product variation create demand for adaptive manipulation across these sectors. Deloitte estimates that U.S. manufacturing could face roughly 1.9 million unfilled positions between 2024 and 2033.

The hardware-agnostic strategy also reduces barriers to international expansion because customers can use locally supported robot brands instead of importing and servicing Phoenix units. Japan is a potential market given the Zeon partnership and the country's aging workforce, while European automotive and industrial customers represent a follow-on opportunity through existing robot OEM channels.

Proof-of-concept gap: Sanctuary's strongest public result is a 40-minute, 313-trial wire-insertion test, but translating that result into sustained, multi-shift production uptime across varied parts, contamination, maintenance events, and operator interactions remains an unproven engineering and commercial challenge.

Capital disadvantage: Sanctuary has raised approximately $149M, compared with over $1.75B for Figure, over $935M for Apptronik, and over $641M for Agility, creating a structural gap in compute infrastructure, manufacturing capacity, deployment teams, and data collection that Sanctuary must bridge through capital efficiency and partnerships rather than spending parity.

Software commoditization: Google DeepMind's Gemini Robotics, NVIDIA's expanding robotics stack, and well-funded foundation-model companies such as Physical Intelligence and Skild AI could reduce Carbon to an interchangeable integration layer, particularly if multi-embodiment control models become broadly available and robot OEMs can license them directly without Sanctuary's involvement.