Reconstruct the Real World with SuperSim

Massive teams. Months of testing. Expensive simulations created from scratch. This is the reality of robotics training today. 

The problem with current-day simulators

Source: Udacity

Beyond labor intensive training in the physical world, robotics–such as autonomous vehicles, drones, or robots–rely on simulation environments to train for different scenarios so that it knows to behave appropriately. Currently, simulations are built with rendering engines, using hand-crafted meshes, materials, and logic. They’ve helped robotics research move forward—but they come with major limitations:

• Domain gap: You’re training your robot with real world data, but your simulator looks like a game, leading to performance reliability issues in the real world which can be catastrophic.
• Time & Cost: Building a simulation often requires a small army of 3D artists, engineers, and integrators, driving up costs and increasing timelines.
• Compute cost: Even once built, high-quality and photorealistic simulation is expensive to run.
• Use your data at scale: Autonomous systems are constantly collecting rich data—LIDAR, video, position data—but almost none of it ever gets brought back into your system.

This is a backwards way of working.

A better way: Re-simulation from reality

This is why we founded Third Dimension AI. At Third Dimension, we’re unlocking possibilities with spatial generation, starting with a new kind of simulator – SuperSim. Instead of constructing synthetic worlds using rendering engines, SuperSim reconstructs the real world in high-fidelity directly from input data, with novel views and 3D object insertion.

Using AI and advanced 3D techniques like neural rendering and Gaussian Splatting, we turn sensor data into dynamic, hyperrealistic 3D environments, without needing a large team of 3D artists or a rendering engine. With SuperSim, you can:

• Debug failures in exact context
• Re-run and validate your software with different behavior and objects
• Test edge cases in a controlled environment
• Build scenario libraries with no artists in the loop
• Create your own repeatable testing scenarios/centers

Integrating SuperSim also allows you to continuously feed the data collected back into your embodied AI system, so you can refine your model and data flywheel in real time. This isn’t only a visual, cost and speed of delivery improvement. It’s a full-stack rethink of how autonomy systems learn, test, and improve.

Where we’re going

Third Dimension is a spatial generation company building fully dynamic3D worlds to power embodied AI systems. Currently we are focusing on robotics, where we can enable fast, cost-effective and photorealistic 3D simulations for autonomous vehicles, drones, robots in industrial and manufacturing environments, and more. 

Beyond robotics simulation, we are also exploring other areas that can be transformed by 3D spatial generation. One such area is entertainment (think games and movies), which has long been limited by the manual and expensive nature of 3D content creation. In the future, we are excited to explore how Third Dimension’s spatial generation technology can enable unprecedented levels of creativity in gaming and movies.

Third Dimension was founded by Tolga Kart and Piotr Sokolski, both veterans in 3D reconstruction and AI. The team is made up of researchers and engineers with experience at Tesla, Wayve, Google, Luma AI, NVIDIA, Niantic, and more. If you are interested in defining the future of 3D, we want to hear from you.