---
title: "Vivarium"
description: "Will Freudenheim, Christina Lu, and Dalena Tran propose Vivarium as shared infrastructure for training embodied AI: intelligence whose capacities depend on a body interacting with an environment."
type: "reading-notes"
authors:
  - "Will Freudenheim"
  - "Christina Lu"
  - "Dalena Tran"
published: "2023"
original_url: "https://vivarium.host/docs/Vivarium_300823.pdf"
canonical_url: "https://antikythera.wiki/work/studio/vivarium"
md_url: "https://antikythera.wiki/md/work/studio/vivarium"
last_updated: "2026-09-06"
site: "Antikythera Wiki"
---

# Vivarium

> Independent, unofficial reading notes from Antikythera Wiki (https://antikythera.wiki). Written with AI from the published text, with citations; not by the work's author and not affiliated with Antikythera (https://antikythera.org). Read the original at the link below.

- **Authors:** Will Freudenheim, Christina Lu, Dalena Tran
- **Published:** 2023
- **Kind:** Studio projects · Studio paper
- **Venue:** Planetary Computation Studio
- **Original:** https://vivarium.host/docs/Vivarium_300823.pdf
- **Length:** 1.9k words
- **This page:** https://antikythera.wiki/work/studio/vivarium
- **Notes generated:** 2026-09-06

## Summary

[Will Freudenheim](https://antikythera.wiki/people/will-freudenheim), [Christina Lu](https://antikythera.wiki/people/christina-lu), and [Dalena Tran](https://antikythera.wiki/people/dalena-tran) propose Vivarium as shared infrastructure for training embodied AI: intelligence whose capacities depend on a body interacting with an environment. Their 2023 studio paper combines a simulation engine, an asset store, and a library of playable training worlds. Human participation would generate diverse experiences, while interoperable components would let researchers accumulate skills across otherwise isolated experiments. A matrix of zero, one, or many humans interacting with zero, one, or many AI agents organizes the proposal. Its six AI configurations progress from individual movement to cooperation among artificial agents and increasingly complex human–AI combinations. The authors envisage these combinations becoming distributed bodies whose intelligence crosses the boundaries of individual organisms. The paper presents a research proposition; it supplies no implementation results demonstrating that the proposed platform closes the gap between simulation and physical reality.

## The argument

The authors begin with a limitation of AI trained primarily on static information. Understanding descriptions of the world does not provide all the capacities needed to perceive, move through, and change it. Embodied intelligence develops through interaction: an agent’s sensors, physical form, and surroundings help determine what it can learn and do. The proposal concerns both individual robots and combinations of humans and machines whose abilities emerge through their relations.

Training such systems directly in physical environments is costly and difficult to repeat under controlled conditions. Toy worlds reduce those environments to bounded simulations, allowing experiments to run repeatedly and generate behavioral data. However, success inside a simulation does not guarantee success outside it. The [Sim-to-Real Gap](https://antikythera.wiki/terms/sim-to-real-gap) arises when simulated physics or circumstances fail to capture what the agent will encounter. The authors argue that diverse shared experience would improve generalization, but that incompatible libraries and isolated research prevent skills from accumulating.

Vivarium addresses that fragmentation through a common platform. Its engine would simulate physical properties beyond visual appearance; its asset store would distribute objects, sensors, skills, tasks, and benchmarks; its world library would recruit participants through play. The comparison with the internet’s contribution to language-model training is specific: a large shared repository makes a different kind of machine learning possible. Here the proposed resource is interactive experience, generated by people constructing worlds and participating in them.

The matrix then turns this infrastructure into a program of capability development. Individual agents acquire motor skills, networks of agents coordinate, and human–AI combinations learn to distribute perception and action. Increasing the number of participants changes the problem rather than merely enlarging it. One human directing many agents requires different abilities from one AI coordinating many humans. Human executive control is itself a variable: some worlds would make the human an appendage of an AI-directed system.

The concluding claim is that these combinations could become nested forms of embodied cognition. A larger agent could contain smaller intelligent agents as functional organs, combining biological and artificial capacities across scales. This is the paper’s speculative destination. The platform description does not establish how skills become portable between different bodies, how interoperability is enforced, or how performance in simulation is validated in physical settings. Those missing mechanisms matter because the argument depends on collective training producing transferable capacities.

## Section by section

### Preface

The preface connects embodied intelligence, simulation, and collective participation. It introduces Vivarium’s three components and treats games as environments for producing skills ranging from movement to coordination. The aim is a world altered by artificial bodies with sensory capacities different from human ones.

### Theoretical Setup

The authors contrast knowledge acquired from static data with learning through physical interaction. Simulated reinforcement learning, where agents improve through feedback on their actions, offers a controllable training setting. They then identify two obstacles: transfer from simulation is uncertain, and research tools are insufficiently compatible. Their response links a technical problem of generalization to an organizational problem of sharing environments, data, and capabilities.

### Introducing Vivarium

The simulation engine would model properties such as heat retention, pressure, and chemical interaction. The asset store would supply both environmental components and parts of agents, including sensors and learned skills. The public world library would bring together hobbyists, game studios, researchers, and players. Entertainment supplies a proposed reason to participate in the production of training experience.

### Scaffolding Physicalized AI

Figure 1.1 crosses three human counts with three AI counts. Its column without AI contains cosmological simulations, single-player games, and multiplayer games. The remaining six cells name distinct arrangements for training cognition. The counts describe participants within a world: a sensorimotor suite can contain no human player even though people design it and assemble the agent.

### Sensorimotor Suite

Single agents practice identifying objects, moving, and handling their environments in accelerated training loops. Players can vary bodies and sensors, while resource constraints encourage forms suited to particular tasks. Flying or amphibious challenges make bodily adaptation part of the problem. A minimum viable sensorium is the set of sensory capacities sufficient for a particular niche, rather than a copy of human perception.

### CogNets (Cognitive Networks)

Multiple artificial agents learn to coordinate through tasks such as monitoring endangered species or assembling machinery. The authors propose that agents could exchange internal states or share sensory access, weakening the boundaries between them. Different sensors create different [Umwelts](https://antikythera.wiki/terms/umwelt), the perceptual worlds available to particular agents. Some simulations prohibit direct communication and require stigmergy: coordination through changes left in the environment. Networks that function as larger organisms introduce the possibility of agents composed of other agents.

### Centaur Syzygy

One human and one AI form a close pairing. They can divide tasks according to their capacities, share control of an exoskeleton, or communicate through remote sensing. The human’s input is responsive but noisy. Some worlds reverse the expected hierarchy by giving the AI stronger sensing and decision-making abilities while the human provides physical action. The pairing therefore makes the allocation of control something to rehearse.

### Omni-Chimera

One human coordinates many AI agents, combining the preceding capacities for inter-agent cooperation and human–AI interaction. Operating a large vehicle or performing surgery illustrates tasks requiring different skills to work together. As the agents learn, human instructions could become more abstract, leaving detailed execution to the group.

### Wetware Choreography

One AI coordinates many humans. It must reconcile competing inputs and direct collective action while the humans also learn to cooperate. A handful of participants creates a problem of integrating discordant commands; hundreds create a problem of recognizing larger patterns of social behavior. “Wetware” names the biological participants from the perspective of the coordinating system.

### Theoretical Implications

The final cell contains many humans and many AI agents in massively multi-agent roleplaying games, or MMARPGs. The authors treat this as the culmination of capabilities developed in the preceding configurations. They propose that repeated combinations could yield bodies within bodies, with intelligence distributed across biological and artificial participants. Their conclusion moves from a training environment to a planetary prospect: the forms that sense and reshape the world need no longer take the individual human body as their template.

## Key concepts

- **Embodied intelligence** — Cognition in which bodily form and interaction with an environment help determine an agent’s capacities.
- **Toy worlds** — Bounded simulations that isolate relevant conditions for training agents and recording their behavior.
- **[Sim-to-Real Gap](https://antikythera.wiki/terms/sim-to-real-gap)** — The failure of capacities acquired in simulation to transfer reliably to physical circumstances that the simulation misrepresents or omits.
- **Capability scaffolding** — Building complex coordination on acquired motor skills and simpler combinations of participants.
- **CogNets** — Networks of artificial agents that share or coordinate cognition and may function as larger composite agents.
- **Stigmergy** — Coordination mediated by changes to a shared environment, used here when direct communication is unavailable.
- **Human–AI configurations** — The matrix’s arrangements of participant numbers, which change how sensing, commands, action, and control must be distributed.

## Connections

- [After Alignment](https://antikythera.wiki/work/journal/afteralignment) explicitly presents Vivarium in “Vivarium in Toy Worlds” and connects it to recursive simulations that intervene in what they model. The paper develops the platform and participant configurations behind Bratton’s brief presentation.
- [Synthetic Counteradaptation](https://antikythera.wiki/work/journal/coginfra/synthetic-counteradaptation) cites Vivarium’s matrix when organizing its speculative scenarios by the number of human and AI participants. It extends the concern with multi-agent interaction toward successive adaptations, deception, and counterdeception; Vivarium places greater emphasis on shared training and composite embodiment.
- [57 Ideas & Questions about Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra-film) revisits Vivarium as a first-year studio project informing later exercises. Its account of simulated interaction and uncertain transfer to the real carries the proposal into the discussion of intelligent infrastructures.


## Terms used

- [Sim-to-Real Gap](https://antikythera.wiki/terms/sim-to-real-gap)

## Related works

- [After Alignment](https://antikythera.wiki/work/journal/afteralignment)
- [Synthetic Counteradaptation: A Principle of Human–AI Coevolution](https://antikythera.wiki/work/journal/coginfra/synthetic-counteradaptation)
- [57 Ideas & Questions about Cognitive Infrastructures](https://antikythera.wiki/work/journal/coginfra-film)