Lab · live demo
Imagination vs reality.
A model-based RL agent plans by dreaming trajectories inside a learned world model rather than acting in the real environment. The catch: the model is never perfect, and error compounds every imagined step. Step the rollout forward and watch the plan peel away from reality.
// LATENT ROLLOUT · PLAN vs REALITY · STEP 28 / 28error 1.40
reality imagined plan
prediction error vs step
The agent plans by imagining forward inside its learned world model instead of touching the real environment. But the model is never exact, and each predicted step is fed back into the next — so a tiny per-step error compounds. Reality stays on its orbit; the plan spirals off. Here the plan is reliable for about 14 steps before error crosses the trust limit. That ceiling is why Dreamer-style agents imagine in short bursts and keep re-grounding in real experience.