Gizmo generates editable 3D environments for robot training from text and images

Why it matters

Gizmo turns simulation authoring into an API call, giving robotics teams a possible way to generate and refresh training environments without rebuilding each scene by hand.

Shrey Kothari (@Shreyko) opened the public beta of Gizmo on July 21st, giving robotics teams an AI agent that converts text and reference images into structured 3D environments they can edit, test and export to common simulation engines.

Kothari introduced Gizmo in a seven-post thread on X, positioning the product as Antim Labs' "world authoring layer for simulations." He founded Antim Labs with Gokul Srinivasan (@gokul8967) and Viswajit Nair (@badbotvivi). The founders previously built under the 4Wall AI name, working on virtual worlds and interactive environments for training and evaluating AI agents. Antim Labs' current site identifies 4Wall AI Inc. as the legal entity behind the product. (antimlabs.com)

Gizmo gives Antim Labs a direct entry into robotics tooling. The founders are betting that simulation creation can become an agent-driven software workflow rather than a manual 3D production job, allowing researchers to generate more environments for training, evaluation and edge-case testing.

Structured scenes instead of opaque meshes

Gizmo accepts a written prompt, as many as three reference-image URLs through its API, or both. Antim Labs says the agent identifies the required objects, generates their components in parallel, constructs a floor plan, places assets using spatial constraints and runs physics and connectivity checks before saving the result.

The technical distinction is how Gizmo represents those objects. Many generative 3D systems produce a finished polygon mesh that is difficult to change once created. Gizmo instead builds structured assets from geometric primitives and records dimensions, transforms, materials, joints, collision settings and other physics properties as editable parameters.

That structure lets a user resize a cabinet, deepen a drawer or change a hinge without regenerating the entire object. Gizmo's editor also supports follow-up instructions such as adding shelves or scattering objects across an existing scene. Users can inspect and manually change the underlying primitives, materials, joint limits, mass, friction and collision approximations. (docs.gizmo.antimlabs.com)

The output is designed for robot interaction rather than visual presentation alone. According to Gizmo's documentation, generated scenes can include real-world scale, semantic regions, sensor anchors, articulated joints, affordance labels, inertial properties and collision geometry. A browser physics preview lets users test whether doors swing correctly, drawers move along the intended axis and articulated components stay attached before export. Antim Labs warns that Gizmo remains in beta and that generations can fail, with regeneration offered as the current remedy. (docs.gizmo.antimlabs.com)

One scene, three robotics stacks

Gizmo exports the same scene to USD or USDZ for Nvidia Isaac Sim and Omniverse, MJCF for Google DeepMind's MuJoCo, and SDF for Gazebo. Antim Labs says the simulator-specific exports preserve joints, masses, inertias, materials and collision shapes. GLB export is also available for visual interchange, although that format drops physics and joint data.

The cross-simulator approach addresses a practical source of duplicated work. A robotics team can author one environment, load it into several physics engines and compare contact, articulation and policy behavior without maintaining a separate version for each stack. Any edits made inside Gizmo flow into the next set of exports. (docs.gizmo.antimlabs.com)

Antim Labs also provides a catalog that it says contains more than 3,000 physics-ready assets, alongside tools for importing USD, USDZ and GLB files and embedding robots during export. The built-in robot library draws from MuJoCo Menagerie and includes robot arms, quadrupeds and humanoids, according to the documentation. The 3,000-asset figure is an Antim Labs count and has not been independently audited. (docs.gizmo.antimlabs.com)

The API is the larger product bet

The browser editor is only one route into Gizmo. Antim Labs has published a REST API that can generate scenes and individual assets, apply natural-language edits, monitor asynchronous jobs and export completed environments without opening the web interface.

Antim Labs is explicitly designing that API for coding agents and automated robotics pipelines. Its documentation describes integrations in which Claude Code, Codex or another tool-using agent detects a missing simulation asset, requests it from Gizmo, waits for generation and adds the exported package to a repository or test run. Teams could also generate a different environment for each continuous-integration scenario or rebuild a digital twin when new reference images arrive. (docs.gizmo.antimlabs.com)

Generation is measured in minutes. Antim Labs estimates that structured or articulated objects generally take about 15 to 20 minutes, while organic objects built through a mesh pipeline take about two to four minutes. Those are product estimates rather than independently benchmarked completion times, and the API returns a job-specific estimate when a request begins.

Kothari said Antim Labs plans to add scene reconstruction from monocular video with articulated assets over the next several weeks. The current public beta already lists video and scans as inputs for Antim Labs' broader real-to-simulation service, while Gizmo's public workflow and API documentation currently center on prompts and reference images.

Gizmo leaves the decisive question to robotics teams: whether its generated geometry and physics survive contact with real training pipelines. Editable structure and multi-engine export give Antim Labs a concrete test. Visual similarity is insufficient for robotics; joints, collisions, dimensions and physical behavior have to remain useful after the scene leaves Gizmo's browser.