Robotics teams now train robots in simulation long before a real machine touches a real factory floor. That work is impressive, but it is also hard to show. A reinforcement-learning policy running in a grey test scene does not convince an investor, a customer or a keynote audience. Somebody has to turn simulation data into a scene people can understand.
That is the work I do at Cineshare. Here is what I learned from one project that went all the way to the NVIDIA GTC 2026 keynote, and how those lessons led me to build Hermes, a supervised 3D workflow for preparing scene content.
The Persona AI project
Persona AI builds humanoid robots for heavy industrial work, such as welding in shipyards. For GTC 2026 they wanted to show their welding humanoid working inside an HD Hyundai shipyard scene.
It is important to be precise about who did what. Persona AI developed the robotics motion pipeline: they captured human motion with Xsens suits and trained a reinforcement-learning policy in NVIDIA Isaac Lab. My part was scene composition, animation integration and cinematic visual production in NVIDIA Omniverse:
- converting the motion data to USD and applying the RL animation to the robot skeletons
- rebuilding materials so the robots and the shipyard held up under path tracing
- placing three humanoids in the shipyard scene
- adding welding spark effects that sit naturally in the light of the scene
We had about one month. The sequence ran in Jensen Huang’s keynote, which Dell reports was seen by an audience of about 20,000 people, with 130,000+ impressions reported separately.
Three lessons from the project
Scene preparation is the real bottleneck. The robot motion was ready. What took time was everything around it: converting assets, fixing scale and hierarchy, rebuilding materials and assembling a scene that renders cleanly.
Be clear about what a visual proves. A cinematic sequence communicates an approved concept. It does not validate robot physics or safety.
Local GPU power changes how you iterate. I rendered the project on a Dell Pro Max Tower T2 with an NVIDIA RTX PRO 6000 Blackwell GPU, reviewing path-traced frames at my own desk.
Hermes: a supervised 3D workflow
Hermes is a Cineshare workflow, still in development, that prepares and assembles 3D content for OpenUSD scenes, with a person reviewing each approval step. My current demonstration uses an industrial rack: Hermes prepares the parts, keeps components a reviewer has approved, builds a revised assembly and renders views for review.
I deliberately do not call Hermes an autonomous agent. Its output is not mechanically validated and is not ready for robot training on its own. Today I use it inside defined pilot projects. Read more about Hermes.
