Simulation Performance

Simulation Performance#

The performance of the simulation can be affected by various factors, including the number of objects in the scene, the complexity of the physics simulation, and the hardware being used. Here are some tips to improve performance:

  1. Use Headless Mode: Running the simulation in headless mode can significantly improve performance, especially when rendering is not required. You can enable headless mode by using the --headless flag when running the simulator.

  2. Avoid Unnecessary Collisions: If possible, reduce the number of object overlaps to reduce overhead in the simulation. Excessive contacts and collisions in the simulation can be expensive in the collision phase in the simulation.

  3. Use Simplified Physics: Consider using simplified physics collision geometries or lowering simulation fidelity for better performance. This can be done by modifying the assets and adjusting the physics parameters in the simulation configuration.

  4. Use CPU/GPU Simulation: If your scene consists of just a few articulations or rigid bodies, consider using CPU simulation for better performance. For larger scenes, using GPU simulation can significantly improve performance.

Collision Geometries#

Collision geometries are used to define the shape of objects in the simulation for collision detection. Using simplified collision geometries can improve performance and reduce the complexity of the simulation.

For example, if you have a complex mesh, you can create a simplified collision geometry that approximates the shape of the mesh. This can be done in Isaac Sim through the UI by modifying the collision mesh and approximation methods.

Additionally, we can often remove collision geometries on areas of the robot that are not important for training. In the Anymal-C robot, we keep the collision geometries for the kneeds and feet, but remove the collision geometries on other parts of the legs to optimize for performance.

Simpler collision geometries such as primitive shapes like spheres will also yield better performance than complex meshes. For example, an SDF mesh collider will be more expensive than a simple sphere.

Note that cylinder and cone collision geometries have special support for smooth collisions with triangle meshes for better wheeled simulation behavior. This comes at a cost of performance and may not always be desired. To disable this feature, we can set the stage settings --/physics/collisionApproximateCylinders=true and --/physics/collisionApproximateCones=true.

Another item to watch out for in GPU RL workloads is warnings about GPU compatibility of Convex Hull approximated mesh collision geometry. If the input mesh has a high aspect ratio (e.g. a long thin shape), the convex hull approximation may be incompatible with GPU simulation, triggering a CPU fallback that can significantly impact performance.

A CPU-fallback warning looks as follows: [Warning] [omni.physx.cooking.plugin] ConvexMeshCookingTask: failed to cook GPU-compatible mesh, collision detection will fall back to CPU. Collisions with particles and deformables will not work with this mesh.. Suitable workarounds include switching to a bounding cube approximation, or using a static triangle mesh collider if the geometry is not part of a dynamic rigid body.

Additional Performance Guides#