# World Models[#](https://docs.nvidia.com/learning/physical-ai/getting-started-with-isaac-lab/latest/transfer-policies-from-simulation-to-reality/04-bridging-the-gap-real-world-data-integration/04-world-models.html.md#world-models 'Link to this heading')

Another interesting topic with real-to-sim is the use of world models. **World models**, including the approach described by David Ha, learn a representation of an environment and predict how it evolves. A latent representation compresses observations into a vector or tensor; whether it retains the information needed for control must be evaluated.

Most often, world models are trained in a specific domain. However, we can also combine multiple domains by training on real-world and sim data, creating unified understanding of both domains.

This unified representation serves as a common language between the two domains, making it easier for the robot to transfer learned behaviors from simulation to reality.

Once the world model is trained, we can develop a policy that operates on its latent representation or learns from imagined trajectories. Ha demonstrated their approach on the simulated CarRacing and VizDoom benchmarks. Applying a world model to robot transfer also requires validating its predictions and the resulting policy on the target system.

A potential advantage is learning features shared by simulation and reality. Training on both domains can support this goal, but does not guarantee aligned representations or successful transfer. Model errors can accumulate over predicted trajectories, so evaluate both predictive accuracy and task performance.

Tip

Learn more: [World Models by David Ha, Jurgen Schmidhuber](https://arxiv.org/pdf/1803.10122)