Accelerating Physical AI with SimReady & VLA Data | Suntae Kim (NdotLight)
Suntae Kim of NdotLight discusses tools for creating simulation-ready 3D assets and VLA data for physical AI at Humanoids Summit.
In this talk
- 0:38
A pipeline for 3D assets
Kim describes NdotLight’s 3D design tools and a database infrastructure solution for creating data. The pipeline takes text or image prompts, generates 3D shapes, and makes the designs available for robots.
- 1:59
Simulation formats and benchmarks
The tools can produce asset details such as key points and segments, and support formats including MJCF and OpenUSD for use with Omniverse. Kim also mentions benchmarking 3D environments.
- 3:11
Assets for varied environments
NdotLight is preparing a project of 3D assets that Kim says is planned for release by year end. The assets can support building environments such as warehouses, logistics spaces, and kitchens.
- 3:41
VLA data and scene variation
Kim describes a VLA data resource and a workflow for setting up environments, poses, and augmented cases. The data can include RGB images, segmentation, bounding boxes, joint matrices, and descriptions tied to specific frames.
- 6:05
Asset preparation takes time
Asked how long it takes to prepare a 3D model for simulation, Kim says the duration depends on complexity and can be about a month. He gives a pallet project as an example and says a fresh food product task took around 30 days.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.