Learn a useful representation.
A predictive model need not reproduce every pixel of a future observation. A joint-embedding predictive approach can learn representations and predict a target representation in latent space.
For a possible gridworld world model, an encoder would map a state to a latent vector. A predictor could use that vector and an action to estimate the representation of the next state, compared against a target-encoder branch.
This is a research direction, separate from the upcoming causal vs full attention work. No implemented JEPA system, experimental result, or academic supervision is claimed.
Start with a world you can inspect.
A small environment makes transitions, obstacles, and the consequences of actions explicit. In the example below, you control the agent directly. It is a deterministic environment simulation, not a learned policy or JEPA experiment.
Navigate to the goal. You control the agent.
Deterministic 4 × 4 environment · 0 actions · no trained Q-table
Questions for an eventual implementation.
- What information should the latent state preserve about the environment?
- How can prediction remain useful without representation collapse?
- How does multi-step prediction error change with the planning horizon?
- How should a latent predictor be compared with a simple transition baseline?
- Which evaluation tasks reveal useful representation learning rather than superficial reconstruction?