Imagine a robot learning to pick up a mug in a virtual kitchen. The mug looks convincing: glossy ceramic, a curved handle, even a small coffee stain. The robot moves a finger toward the handle’s opening—and hits an invisible wall.
In this invented example, the problem is the mug’s hidden contact shape. Its appearance includes a hole, but the simulator treats the whole object as a solid box. A beautiful picture has become a misleading practice object.
That is the problem behind NVIDIA’s October 21 announcement of SimReady Central, a freely accessible library of 3D assets for simulation. The useful idea is straightforward: objects for robot training need information about how they behave, as well as how they look.
The picture and the object you touch
A simulator can use one shape to draw an object and another to calculate contact. The second is called collision geometry. NVIDIA’s collider documentation describes several representations, including simplified meshes and bounding boxes. These approximations can reduce the work needed to calculate collisions, but the choice matters for the task.
A rough box might be adequate when arranging sealed cartons on a pallet. It would be a poor choice for our mug if the robot needs to put a finger through the handle. The relevant detail is the space the finger must pass through.
Shape is only part of the description. How much does the mug weigh? How readily does it slide against the gripper? The OpenUSD physics specification separates mass properties, collision shapes and physics materials, which describe properties such as friction and bounce.
Give the simulated contact too much friction and a weak grip might appear sufficient. Change the mass and the effort needed to move the mug changes. A robot’s practice environment can reward a strategy that depends on those settings.
What the new library supplies
SimReady Central supplies assets in OpenUSD, a format for describing 3D scenes. It is developer material for assembling simulations; downloading a mug does not give you a trained robot.
The important detail is each asset’s declared capabilities. SimReady uses profiles to group requirements for different uses. Its Robotics Prop profile, for example, distinguishes a required core from optional features. A “simulation ready” label is therefore a starting point: check what the asset includes and what your chosen simulator and task require.
Does it have the data—or does it work?
NVIDIA’s validation guide makes this distinction explicit. Schema validation checks requirements without running a physics engine. Behavioral benchmarks open the asset in an engine and exercise it. Depending on the asset’s declared features, those tests can include dropping it, letting it settle, or grasping and lifting it.
Passing a data check does not establish that our mug behaves convincingly. Passing a simulated grasp test still leaves a further question: does a real mug respond similarly? Neither check alone proves that a trained robot will succeed outside the simulator.
For someone watching a robot-training demo, this offers a useful place to look. Notice the object’s behavior: where contact happens, whether it slips, and how it moves when lifted. The coffee stain makes the scene believable. The less visible details make it useful practice.