From a 5-Minute Walk to a Robot-Ready Digital Twin

Santana Row in San Jose is one of my favourite scenes — a 3D Gaussian splat created by Splatica’s fully automated pipeline, which I filmed right after the Humanoids Summit in December. It’s ready to import into NVIDIA Omniverse or Isaac Sim, and it came from a simple 5-minute walk-around.

The idea is simple: we can create a 3D reconstruction of any place — indoors or outdoors — with a consumer-grade Insta360 camera you can buy in any electronics shop. So what’s the real-life usage? It comes in four layers.

Layer 1 — Photorealistic representation

A photorealistic representation of the space, with accurate measurements. Set up virtual cameras inside the sim and you can collect visual data for:

• Training visual models
• Navigation and perception
• Marketing and demos

Layer 2 — Collision meshes

We build collision meshes from the 3D point cloud, so any system inside the simulation understands the boundaries — floors, walls, objects. Great for:

• Collecting hundreds of hours of data
• Fine-tuning walking and locomotion
• Navigation and obstacle avoidance

Layer 3 — Segmented objects with physics

We segment the scene and extract separate objects, using generative AI to create shapes for objects we’ve never seen before — converting them into splats / USDZ so they can be reimported and manipulated (pick-and-place, lifting). Every extracted object gets a weight, material, and other physical properties, so it’s accurately represented in the sim.

It’s not just a visual copy anymore. From a 3D point cloud we now have a representative copy of the physical space — where robots can live, learn, and train 100× faster than in real life.

Layer 4 — Dynamic, interactive objects

Finally, we replace known objects with dynamic objects that have accurate physics — doors, handles, buttons, latches. All the complexity of the physical world, ready for manipulation. Good for:

• Teleoperation inside the sim
• Agentic, prompted data collection across thousands of backdrops, scenes, and lighting conditions

Why it matters

I believe that by creating accurate simulation we can speed up the development of physical AI and the deployment of robotics. Cost-effective data collection and asset creation is one of the keys to doing it at scale.