Limitations

Sensors support

Gz sim supports a wide variety of sensors, like cameras, lidars, … However, ScenarI/O and consequently gym-gz are not yet able to extract data from sensors. Follow #199 if you’re interested in sensors support.

Performance

When operating on a single model, DART provides good enough performance and accuracy for most of the use-cases. However, we noticed that it does not perform well when many models are inserted in the simulated world, especially in contact-rich scenarios.

If your aim is performing simulations for robot learning, we recommend running multiple instances of the simulator, each of them operating on a world containing a single robot.

If your aim is simulating a big world where the controlled model can move inside the scene, exploiting a new feature of Gz sim called levels could be the proper solution. This feature is exploited by the big worlds of subt.

Instead, if your world has many models and the usage of levels does not apply to your use-case, you can try to switch the physics backend to an engine that handles better your simulation.

Note

At the time of writing only DART is officially supported. There is some recent activity to implement the bullet physics engine, but this back-and is not yet ready. As last resort, you can implement a new physics backend following the instructions.