NVIDIA Research treats training as skill refinement instead of gradient descent, so the result is a repo of robot skills rather than model weights. The loop ships as skills for Claude Code and Codex.
For robotics teams: build a skill library you can read and reuse instead of reprogramming each cell from scratch.
Upload a LiDAR or SLAM scan from any scanner, up to 200 GB per file, and get back a twin with every object separated and editable and robots such as autonomous forklifts, pallet AMRs and palletizing arms placed on your floor. You can try layouts, fleet sizes and routes before a robot arrives.
For plant, warehouse and yard managers planning automation: test the robot layout on your real floor before you buy.
Ask in plain words for selection sets, a hard clash run between mechanical and structural, and only the new clashes.
For VDC and BIM coordinators: skip the clicking and get the clash run you asked for.
His skill wasn't code, it was running the agent like a job: break the work into tasks, set limits, and demand tests and proof. People in the trades already manage work that way.