Vision-Language-Action Agent with Latent World Modeling
Tutorial building an embodied agent that learns perception, planning, and replanning from raw pixels using a lightweight world model and model predictive control.
The agent operates in a NumPy-rendered grid world, observing RGB frames instead of symbolic state—closer to real-world constraints than traditional approaches.
Useful for anyone exploring embodied AI, sim-to-real pipelines, or understanding how to combine vision encoders with planning under latency.