Why we rebuilt Rigyd around an agent that reasons first
Ugur Yekta Basak · Published August 21, 2026
Nobody ever figures out what life is all about, and it doesn’t matter. Explore the world. Nearly everything is really interesting if you go into it deeply enough.
Where we come from
We are robot learning and spatial computing engineers who spent seven years building 3D products together. We trained pipelines for AI-based asset creation, fine-tuning large models on proprietary data to cut hallucinations and produce the most realistic results for specific use cases. We optimized assets for real-time rendering on ordinary hardware. We built the SDKs and APIs that delivered tens of millions of immersive experiences.
Then something clicked. Robotics companies started reaching out and asking for help creating simulation scenes.
What we heard
The teams were different in every way except one. Some followed a sim-first approach for training. Some used simulation for evaluation only. Some just needed to test before real deployment. Every one of them needed a different type of solution, and every one of them had simulation engineers spending their time hand-authoring scenes: days per object, weeks per environment. They needed a reliable pipeline for scaling simulations, and they needed both things at once — visual fidelity where the camera looks and physics accuracy where the robot touches.
We launched our v0 three months ago, and we were lucky to see interest from robotics teams across warehouse automation, data center operations, bio-manufacturing, humanoids, industrial automation, and frontier labs.
What three months taught us
After thousands of generations and deep conversations with some of the sharpest engineers in robotics, the lesson was clear: every engineer has a different use case, and a fixed, top-down generation pipeline was not enough. It breaks exactly where robotics gets interesting — dexterous manipulation, objects where articulation is essential, deformables that ignore rigid-body assumptions.
The same object should be built differently for pick-and-place than for dexterous manipulation, because the task decides which details must be exact and which can be simplified away. No fixed pipeline knows that. Something has to reason about it first.
So we rebuilt around reasoning
Version 1 is an agentic system, and the order it works in is the product:
- 1Intake — what does the robot need to touch, and what is the policy learning? Scope first, because it changes what has to be articulated.
- 2Research — real dimensions, material, mechanism. The trace is yours to read.
- 3Composition — parts, per-part physics, joints. Nothing is a single welded mesh.
- 4Approval — you see the composed object before the compute is committed.
- 5Validation — drop and settle tests with automatic correction, then SimReady conformance scored per profile.
- 6Export — OpenUSD and MJCF with the values written in, dropping straight into Isaac Sim, MuJoCo, and more.
On standards
Validation should not be a vendor’s private opinion. NVIDIA’s
SimReady
Foundation defines open, testable specifications for simulation-ready OpenUSD content:
single requirements grouped into features, features bundled into profiles like
Prop-Robotics-Neutral and Prop-Robotics-Physx that act as contracts
between asset creators and consumers. We score our assets against these profiles, and the
validation report travels with every asset.
Why this matters
Billions of robots are coming, and they need endless variety: warehouses, factories, kitchens, and countless ways of interacting with the objects inside them. Real-world data does not scale — one hour of real-world data costs one hour of human labor. The engines have matured. The content is the bottleneck we are removing.
Try it, and tell us what you are simulating. We read everything.