Vivarium is an ecosystem of “toy worlds”: gamified simulations in which different human-AI configurations scaffold complex embodied intelligence.

Embodied intelligence refers to types of cognition in which the “body” plays a decisive role. Artificializing such processes is necessary for AI to make the leap from language-wielding to world-making. Yet training in the real world is prohibitively expensive, impossible to parallelize, and difficult to control. Instead, agents are trained in toy worlds: controlled simulations in which reality is abstracted into a bounded domain. Toy worlds allow for quick iteration, broad proliferation, and streamlined data synthesis in pursuit of embodied intelligence. Negotiating the intricacies of human interaction also becomes possible at scale within them.


If AI agents are to exist in the real world, they will require a collaborative repository and training platform: Vivarium, a diverse ecosystem of interoperable toy worlds. Composed of an asset store, a simulation engine, and a toy world library, Vivarium turns training embodied agents into play. Within these gamified worlds, different genres of human-AI configurations encourage learning from basic movement to interactive negotiation, adversarial feints to stigmergic coordination. Future agents will have sensoriums as diverse as the physical forms they take; they will rehearse and scaffold complex embodied intelligence across the sim to the real. Through Vivarium, an unrecognizable world remade by physicalized AI becomes possible.

READ ⟶

HUMAN-AI CONFIGURATIONS

NO AI

SINGLE AI

MANY AIs

NO
HUMAN

Sensorimotor Suite scene Sensorimotor Suite
CogNets screencap CogNets

SINGLE
HUMAN

Single Player
Centaur Syzygy screencap Centaur Syzygy
Omni-Chimera screencap Omni-Chimera

MANY
HUMANS

MMORPG
Wetware Choreography screencap Wetware Choreography
MMARPG screencap MMARPG