Sanctuary AI Phoenix

A humanoid built for work

Phoenix is a humanoid general-purpose robot that Sanctuary AI unveiled in May 2023. It is the company's sixth-generation general-purpose robot, and Sanctuary described it as the world's first humanoid general-purpose robot powered by its Carbon AI control system^[1]. The company framed the machine as a robot designed for work^[1].

The sixth-generation label points to a line of earlier machines rather than a first attempt; Phoenix is the version Sanctuary chose to present publicly as a general-purpose platform. "General-purpose" here means the same body and control software are meant to be pointed at many different jobs instead of being built for one fixed task.

Body, payload and hands

Sanctuary's official specifications list a height of 5 ft 7 in, a weight of 155 lbs, a maximum payload of 55 lbs and a maximum speed of 3 miles per hour^[1]. The height puts the robot close to the scale of an adult person, which matters if it is expected to reach benches, shelves and door handles in ordinary workplaces. Payload sets the ceiling on what it can pick up and carry: a 55 lb limit covers many boxes, parts and tools, but not everything a person can lift.

The hands are a central part of the design. Official specifications list robotic hands with 20 degrees of freedom in total, haptic technology mimicking a sense of touch, and fine manipulation^[1][2]. A degree of freedom is one independent axis of movement — a finger that can only bend at one joint has one, and a hand that can also spread its fingers or rotate its thumb has more. Twenty degrees of freedom spread across two hands gives the robot a wide repertoire of grips, from a whole-hand grasp on a tool handle to a pinch on a small component. Haptic technology means the hands are built to register physical contact, so manipulation does not depend on vision alone.

Carbon, the control system

Sanctuary describes Carbon as a cognitive architecture and software platform that translates natural language into action with explainable and auditable reasoning, task and motion plans^[1]. In practice that chain runs from a spoken or written instruction, to a plan of what steps to take, to a plan of how the body should move to carry them out. Calling the reasoning explainable and auditable means the system is meant to produce a record that a person can inspect — what it decided, and on what grounds — rather than an opaque output.

Carbon combines symbolic and logical reasoning with modern LLMs, deep learning and reinforcement learning, and uses photo-realistic and physics-realistic simulations for training^[1]. Those ingredients do different jobs. Symbolic and logical reasoning handles explicit rules, goals and constraints; LLMs (large language models) interpret the instruction and the surrounding context; deep learning and reinforcement learning build the low-level skills that turn intentions into movement. Training in photo-realistic and physics-realistic simulation lets behaviour be practised in software worlds that approximate both the look of real scenes and the way objects behave when pushed, lifted or dropped.

Tasks and recognition

Sanctuary said its technology was already capable of completing hundreds of tasks identified by customers across more than a dozen industries^[1]. That claim is about the breadth of tasks the system had been made to handle, and it is the company's own account of where the platform stood at launch.

Sanctuary also said Phoenix was the only humanoid general-purpose robot on TIME's 2023 Best Inventions list^[2], placing the machine in a wider public conversation about robots intended for real work.

Where Phoenix sits

Phoenix belongs to a generation of humanoid robots aimed at ordinary workplaces rather than staged demonstrations. Two things distinguish the approach. The first is the emphasis on a software layer — Carbon — as the product's core, with the body treated as the machine that executes its plans. The second is the combination of language-level reasoning with hands that can feel what they touch, which is what a robot needs if a person is going to hand it an instruction in words and expect a physical job to come back done. The published specifications describe the hardware; the task breadth and the invention-list recognition are Sanctuary's own statements about the platform.