From photographs to a lab you can explore: building with Astra

An ICI Innolabs experiment in turning 2D references into editable 3D models and an interactive browser experience.

A photograph can show you a laboratory. An interactive model lets you walk between its desks, inspect a robot, move a chair, and see how the space fits together.

Our Lab3D demo explores that transition. Using photographs of our laboratories as references, we created three reconstructed spaces with equipment, animated demonstrations, and objects visitors can interact with in a desktop browser.

Try the live demo in your desktop browser to explore the rooms and see the research scenarios in action.

We credit OpenAI’s GPT-6 Astra for its contribution to creating the models and the demo. The capability we wanted to explore was practical: using visual references and instructions to help produce editable 3D scenes through code and modeling tools. Astra’s contribution sits within that workflow, alongside our reference material and the existing hardware designs credited below. OpenAI describes Astra as a model for complex reasoning, coding, and computer use.

Astra also had access to a selection of our research results. We tasked it with imagining and implementing interactive scenarios that would highlight four promising results, making their potential easier for visitors to understand. This extended its role beyond reconstructing the laboratories to deciding how research could be presented through actions, animations, and interactions within them.

Three rooms, three ways to explore

The first room introduces the AI lab. Workstations and everyday furniture establish the space, while a reconstructed Robot Nano Hand brings it to life. The hand cycles through an open palm, a fist, an OK sign, a peace sign, and a pointing gesture. Visitors can open a close-up, rotate their view, inspect animated joint angles, and pause or restart the sequence.

The second room puts robotics on the workbench. OpenBot Waffle follows a predefined route around the central bench, stopping when moved furniture blocks its path. Its inspection panel includes an orbitable model and a simulated LiDAR view. Nearby, a SO-ARM101 arm repeatedly sorts six colored cells into labeled holders, making a complete pick-and-place sequence easy to watch from different angles.

The third room reconstructs the VR lab from a panoramic photograph. Desks, headsets, equipment trolleys, and a treadmill with a harness give the room its character. Selecting the treadmill opens VR Rush, a browser game where visitors collect pickups, jump barriers, and slide beneath gates. It runs with a keyboard and mouse; no headset is required.

Across the rooms, visitors can rearrange selected furniture and monitors, inspect equipment, and switch ceiling lights on and off. In the first room, handled windows also tilt inward. These small interactions make the reconstruction a space to investigate.

From 2D references to editable 3D assets

The source photographs provide visible details: the arrangement of desks, the shape of equipment, the color of blinds, and the clutter that makes a working laboratory recognizable. Turning those observations into a navigable scene also requires decisions about depth, scale, hidden surfaces, and how objects fit together.

The project preserves those decisions in editable Blender scenes and reusable asset libraries. Room geometry is exported to GLB for the browser, where Three.js renders the environment and SvelteKit provides the application interface. The robot hand and sorting arm retain their animations through the export process.

That editability matters. A desk can be repositioned, a material corrected, or a robot inspected as separate parts. The result gives us something we can refine and reuse for future demonstrations.

It also makes the Astra contribution concrete: AI assistance can support the path from visual interpretation to geometry, animation, and interactive software. This demo is a project example of that workflow, rather than a benchmark of reconstruction accuracy.

A visual reconstruction with clear limits

Room dimensions and many equipment details are estimates from photographs. Hidden surfaces are inferred, and collision shapes are simplified for walking and arranging objects.

The robot behaviors are demonstrations. OpenBot follows a scripted patrol, its radar is simulated, and the arm and hand play prepared animations. The browser does not control physical hardware or receive live sensor data.

Those boundaries make the demo useful for its intended purpose: letting people explore the spaces and understand the equipment through interaction.

Why we built it

At ICI Innolabs, our work connects artificial intelligence, robotics, and immersive media. This experiment brings those interests together in a format that people can explore directly.

A virtual lab could help introduce a research space before a visit, explain a mechanism from several viewpoints, or provide a starting point for an educational demonstration. Here, we have a concrete foundation: three rooms, editable models, and interactions that make the equipment easier to examine.

Credits

AI-assisted creation: OpenAI’s GPT-6 Astra, credited for model creation, development of the interactive demo, and imagining and implementing scenarios to highlight four promising research results.
Research results, project context, and laboratory references: ICI Innolabs.
Modeling and browser tools: Blender, Three.js, and SvelteKit.
OpenBot Waffle: Original printable CAD from ICI-Innolabs/OpenBot-Waffle, used under the MIT license and supplemented with photo-informed details.
SO-ARM101: Original SO-101 meshes and robot description from TheRobotStudio/SO-ARM100, used under Apache-2.0 and adapted to the photographed setup.
Robot Nano Hand: Locally reconstructed geometry informed by our photographs, with mechanism references credited to Robot Nano Hand—Rob Knight / The Robot Studio, Silicon Highway, and Codethink—and the gesture sequence informed by Hamsa.

Explore the live demo and take a look around our virtual laboratories. Curious about applying this approach to a research space or an educational experience? Get in touch with ICI Innolabs.