BlogJune 2, 2026

Scan to SimReady: How Lyte Closes the Loop from Sensor to USD

Alexander Shpunt

CEO & Co-Founder, Lyte AI

Lyte's perception solutions are in active deployment with leading robotics OEMs. Today we're sharing the second half of that story: Lyte enables perception-to-twin (scan-to-twin), the layer that turns every scan into an indexed, SimReady asset, and every redeployment into a refresh of the library. Lyte is currently engaged with global industrial leaders in Asia and Europe on perception-to-twin programs, built on the same stack as the perception layer.

Reality capture is the bottleneck holding back industrial digital twins. The hard part is not stitching geometry in a sim engine. The hard part is acquiring a scene with enough fidelity, geometric precision, and semantic consistency that the downstream pipeline has something real to work with. Most capture stacks struggle here. Sensors disagree on time, geometry, and what they are looking at. The compute layer inherits a noisy, fragmented input and tries to clean it up after the fact.

Lyte solves the capture problem at the source. LyteVision combines RGB, 4D imaging, and IMU into a single multimodal sensor. All modalities are synchronized, calibrated, and fused at the silicon level on the Lyte SoC. The output is a structured, high-fidelity capture (at the edge) that operates from bright sunlight to complete darkness and stays robust under optical interference, dust, and weather. This is the foundation everything else stands on.

One pipeline, three markets

The same perception-to-twin loop serves three distinct customer needs. Assets are the unit of value. Recurring scans are the operating model. The Lyte stack compounds underneath the robot and its surrounding infrastructure at every stage.

Robot OEMs need a recurring digital twin for locomotion. Environments change daily; a one-shot scan ages out in weeks. The same sensor the robot uses to perceive at run-time is the sensor that keeps the twin current.

Virtual Facility Integration customers need indexed, SimReady assets using the OpenUSD framework to handle complex industrial scenes — legacy equipment, custom tooling, mixed materials, constant operational change.

Training programs and humanoid labs need arm manipulation grounded in pose estimation. The same asset library answers two questions at once: is this a known object in a new pose, or something new that needs reconstruction; and where exactly is it, in what state. Both answers come from the same library, kept current by recurring scans through the same sensor stack.

How the pipeline works

On top of the capture layer, the infrastructure runs on the NVIDIA CAD2SimReady skill, with output validated against the NVIDIA SimReady framework:

Click to expand

Capture. LyteVision acquires geometry (dense 3D point cloud), dynamics (per-point Doppler velocity), and color (RGB). All modalities synchronized, calibrated, and fused on the Lyte SoC — at the edge. The sensor hardware unifies every input stream before software ever touches it, which means the reconstruction layer always starts from a consistent, latency-aligned multimodal capture. No post-hoc alignment, no temporal drift between sensor modalities.

Click to expand

Reconstruction. Lyte's proprietary reconstruction engine produces candidate objects, runs either on-device at run-time or on-server for post-processing, and outputs each candidate as a USD asset with geometry. Geometric mesh generation and semantic decomposition are accelerated by differentiable CUDA kernels written in NVIDIA Warp.

Run-time optimization. LyteIQ aligns and segments the scan. Semantic logic can refocus LyteIQ at run-time to optimize for segmentation and reconstruction success — extracting collision meshes and preparing the scene for downstream material and semantic assignment.

Material and semantics. Built on NVIDIA Content Agent (part of the CAD2SimReady skill), Lyte's asset generation pipeline adds semantics, metadata, and material and physics properties. Assets are indexed and stored in asset libraries that support USD Search.

Click to expand

Robot run-time reconciliation. At run-time, the robot scans the scene, searches the scene graph, matches observed objects against the asset library, and reports back: known object (use existing asset) or unknown (trigger reconstruction and indexing). This is what makes the asset library compounding rather than static.

Output. SimReady assets conforming to the CAD2SimReady skill, validated by the SimReady framework. Each asset carries geometry, collision mesh, material properties, and semantic tags — ready to drop directly into NVIDIA Isaac Sim.

Each scan produces an OpenUSD scene that flows directly into NVIDIA Isaac Sim, where SimReady assets become the training and synthetic-data ground for robotics policies.

What's next

First SimReady assets from these engagements will be published on lyte.ai. We are extending the pipeline against the requirements of industrial reality capture — larger facilities, more complex geometry, mixed surface materials.

If your team is working on industrial reality capture or contributing to the SimReady standards under the SimReady framework, we'd like to compare notes. Also see NVIDIA's announcement of Isaac GR00T platform updates.

NVIDIA powers compute. Lyte powers reality.

Reach out

info@lyte.ai