Simulation: A Game-Changer#

Fortunately, simulation offers a solution to many of these problems. At NVIDIA, we’ve developed expertise in this area, and it’s become a core part of the Omniverse platform.

Benefits of Simulation#

  • Speed: Simulations can run much faster than real-time. Parallel environments collect experience simultaneously, increasing the aggregate simulated experience per wall-clock second. Throughput depends on hardware, scene complexity, sensors, and environment count.

  • Access to Information: In a simulation, you have access to all the data about the robot and its environment. This “privileged information” can be invaluable for training and debugging.

  • Scenario Generation: You can procedurally generate an infinite number of scenarios, exposing the robot to a wide range of situations it might encounter in the real world.

Tip

Learn more: Learning a State Representation and Navigation in Cluttered and Dynamic Environments

Sim-to-Real Transfer#

This is where the concept of sim-to-real comes into play. The process involves training a policy in simulation, then transferring it to the real robot.

Training in Simulation vs Zero-Shot Deployment

ANYmal parkour: Learning agile navigation for quadrupedal robots. (Published in Science Robotics, https://www.science.org/doi/abs/10.1126/scirobotics.adi7566)

In practice, this means taking the weights of the trained neural network and loading that file onto the robot’s hardware for inference–this is how you tell the robot what to do.

However, it’s not quite as simple as it sounds. There’s a catch.