
GENESIS
THE PLATFORM FOR UNIVERSAL OPERATORS
In ancient tales, gods spoke worlds into existence.
Genesis is a platform to generate universal operators, spawning them into reality, animating and managing them across their lifecycle.
A Universal Operator is an AI worker capable of operating in both digital and physical worlds. It performs complex planning, acting in the real world using tools and services, and creating new ones when needed. Operators can also collaborate with or spawn and direct other operators forming dynamic networks or ecosystems of intelligence.
ORCHESTRATING REALITY
To function effectively, universal operators must plan, visualize, predict, and explain their future actions. This ensures alignment with human intent, safety in execution, and clear decision-making. When confronted with a novel problem or a user who is undecided, operators analyze the situation and offer actionable options with clear tradeoffs, costs, and outcomes.

.png)
.png)
Target Image
_gif.gif)
_gif.gif)
_gif.gif)
Predicted Video
Planning is inherently a search problem that scales with task complexity
While simple tasks can be handled by single-shot predictive or autoregressive models, more novel or complex challenges—those outside already-mastered domains—require inference time techniques similar to AlphaGo or the o1 model class. Genesis operators are capable of an efficient form of this high-level planning.
These operators seamlessly integrate existing tools and services from the economy, including APIs, logistics systems, and robotic tools, to achieve their goals. They evaluate tradeoffs and costs to determine efficient paths, enabling them to execute diverse tasks with minimal waste—whether coordinating deliveries, automating factories, or operating entire storefronts.
THE GREAT CHAIN OF MAKING
To act in the world, universal operators utilize tools, which can range from software programs in the digital realm to robots in the physical realm. Our definition of robot encapsulates any controllable tool in the physical world, a recursive definition as a robot can drive other robots.

DYNAMIC TOOL CREATION
Each universal operator functions as both worker and toolsmith. When faced with a task, an operator evaluates whether creating a new tool would be more efficient than using existing ones. This toolsmith capability spans both digital and physical realms. In the digital domain, operators can create new programs and services to extend their capabilities.
In the physical world, embodied operators can design robots from scratch and orchestrate their creation using existing tools like robotic arms and 3D printers. See our papers on Robot-build-robot and Robot designer for early examples of these capabilities in action. This points toward an extraordinary future: one where civilization can generate custom robots of any form on-demand, completely autonomously, through networks of distributed superfactories.
An operator can also spawn other universal operators or more specialized variants with specific capacities and competencies, making them more efficient for certain tasks—effectively creating entire operator corporations that self-organize.
GENESIS
A Living Culture of Skills
Universal operators possess inherited skills derived from shared action models, a culture refined across generations of use in various embodiments. When faced with an entirely novel task outside its current capability, the unskilled operator must learn to master the skill, either through human guidance or autonomously.
SKILLS LEARNING

Human-guided learning allows embodied operators to acquire new skills through natural interface.
"SHOW AND TELL"
Human-guided learning allows embodied operators to acquire new skills through natural interface.
SELF LEARNING
For autonomous learning, operators must learn efficiently by generating their own training data.
In the digital domain, this might involve crawling the internet to gather relevant data.
In the physical realm, it could mean conducting active learning through a small number of safe, intelligent trials.


SELF LEARNING
For autonomous learning, operators must operate efficiently by generating a minimal number of safe trials. In the digital domain, this might involve crawling the internet to gather relevant data. In the physical realm, it could mean conducting active learning through a small number of safe, intelligent trials.
MASTERING MULTIPLE ROBOTIC SKILLS
From human guidance or learning autonomously, our embodied operators have learned multiple useful skills including liquid pouring, powder scooping, cutting, picking, packing, wire handling and plugging and screw tightening as shown below.
