Frontier · 2018– · predictive representation
World Models
Models that learn predictive representations of environments: compressing observations into latent states and modeling their dynamics to imagine futures.
ModLensFrontier
Interactive Diagram
Focus the lens
Click any component to read what it does. Signal direction follows the edges.
inputObservation. Raw experience from the environment.
Core idea
Intelligence that plans needs a simulator. A world model learns the dynamics of an environment in latent space, letting systems rehearse consequences before acting.
Why it exists
Acting in the real world is expensive and risky; systems that can simulate outcomes internally can plan, explore and learn far more safely.
Data Flow
What moves through the system
- 01Observations are encoded into latent states.
- 02A dynamics model predicts future latents under actions.
- 03The system imagines rollouts in latent space.
- 04A policy or planner selects actions using imagined futures.
Strengths
- + Sample-efficient planning
- + Safe rehearsal of consequences
- + A path toward general embodied intelligence
Limitations
- − Latent dynamics drift from reality
- − Evaluation is an open research problem
Applications
- · Robotics
- · Simulation
- · Game agents
- · Long-horizon planning research
