Skip to content
rfranrPublic

About

Explores how AI agents interact with a shared environment through constrained interfaces.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

WorldGL

WorldGL is a local Three.js + TypeScript flight prototype and a small experimental environment for studying agent navigation through structured tools. It includes a browser simulation and a local experiment inspector. The simulation needs no backend or runtime network service; the optional Pi experiment uses your configured model provider.

WorldGL explores how AI agents interact with a shared environment through constrained interfaces. The goal is to experiment with different levels of perception, available tools, and navigation semantics, observing how these constraints influence agent behavior. The world remains the same, but what an agent can perceive and do may differ.

Third-person craft above faceted terrain, with geometric landmarks, player telemetry, recorded paths and a north-up minimap.

WorldGL’s flight environment, with terrain, landmarks and live telemetry.

Implemented today

  • Deterministic low-poly terrain, a geometric craft and three initial landmarks.
  • Desktop keyboard flight, following camera, terrain clearance and live telemetry.
  • Structured inspection with distance and horizontal/vertical field-of-view limits.
  • Asynchronous movement, queued destinations, routes and explicit cancellation.
  • Optional trajectory recording, JSON export, 3D paths and a persistent minimap.
  • Local session/trajectory import, multi-run comparison and evidence pinning.
  • An optional isolated-tool Pi adapter for model-driven exploration.

This is direct velocity control, not a physics simulation. Landmarks have no collision response. There is no world wrapping, imported-model pipeline, persistent world storage or object-editing UI.

Quickstart

Requires Node.js 20.19+ or 22.12+ and a modern WebGL-capable desktop browser. From a checkout:

npm ci
npm run dev

Open the URL printed by Vite (normally http://localhost:5173). Use W/S to fly, A/D to turn, Q/E to strafe, Space/Shift to rise/descend and R to reset. Click back into the scene after using debug controls to resume keyboard flight.

Open the HUD's Inspect recorded experiments link, or /inspector.html, to load local extracted JSON/JSONL artifacts. Imports stay in the browser. Deploy only the built dist/ directory, never the repository root.

Agent-driven navigation

The optional Pi explorer connects your configured model to WorldGL through a restricted set of tools. The agent uses inspect_world for perception-limited observations, requests movement with move_to or execute_route, and reads navigation status to check progress. Tool requests express intended actions; returned outcomes report execution.

Pi terminal with inspect_world and execute_route calls beside WorldGL showing the craft, recorded paths and minimap.

Pi tool calls and returned route outcomes alongside the WorldGL scene. The agent's completion statement is not an independent evaluation of exploration success.

Development checks

npx playwright install chromium # explicit browser provisioning, including CI
npm run check:report            # lint, tests, build, complexity and architecture
npm run preview                 # serve the built production bundle

The suite includes real Chromium interaction regressions and adapter unit tests. Passing tests do not establish agent exploration quality or visual usability. See validation and evidence for the inspected results, limitations and untested hypotheses.

Documentation

For changes, preserve public APIs and follow agent/development guidance. No license file is currently provided in this repository.

About

Explores how AI agents interact with a shared environment through constrained interfaces.

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages