BowerBot

BowerBot is an Apache 2.0 open-source agent that turns natural language prompts into validated, ASWF-structured OpenUSD scenes using existing assets, exporting to .usda or .usdz for Omniverse and Apple Vision Pro.

BowerBot is an open-source AI agent that assembles OpenUSD scenes from natural language descriptions. A user types a prompt such as a modern kitchen with marble countertops and pendant lights, and the agent builds a stage file, resolves matching assets, places them in space, sets up lighting, binds materials and validates the result. The homepage demo shows a kitchen scene finished in 18 seconds with five assets, four lights and two materials, then packaged as a USDZ file. A key design choice is that BowerBot assembles existing assets and never generates geometry. Assets come from sources such as Sketchfab, a local digital asset manager, or a custom plugin the user writes through its skills system. Output follows Academy Software Foundation structure, so the resulting hierarchy of Xforms, scopes, lights and materials is already organized the way a studio pipeline expects. Six light types (dome, sphere, rect, disk, cylinder and distant) are placed automatically, and a built-in validation loop lets the agent check its own work and correct errors until the scene is right. The tool targets people who need USD scenes without hand-authoring every layer: VFX and animation teams bootstrapping scenes, spatial computing developers producing USDZ for Apple Vision Pro, simulation and synthetic data teams that need randomized training environments at dataset scale, and pipeline technical directors who want unit handling and compliant structure automated. Scenes export as .usda or .usdz for Omniverse, Vision Pro or any USD-compatible viewer. In practice BowerBot runs as a command-line tool built on Python 3.12+, OpenUSD 26.x, LiteLLM and Pydantic, with support for Claude and GPT-5 models, and it is released under Apache 2.0. It differs from text-to-3D generators because it composes scenes from real, existing assets rather than inventing meshes, which suits pipelines that already own asset libraries. It is built by Binary Core, a studio working on digital twins and physical AI, and it is a developer-oriented tool rather than a point-and-click design app.