# Ctrl-World-Graph

This directory is a lightweight graph-conditioned fork scaffold for `Ctrl-World`.

Goals:
- Keep the original `/workspace/Ctrl-World` repo untouched.
- Reuse the original diffusion backbone and pipeline where possible.
- Add a new graph-conditioning path built with PyTorch Geometric.

The first implementation target is:

1. Encode each future graph frame with a PyG spatial encoder.
2. Compress variable-length node embeddings into a fixed number of graph tokens.
3. Model graph-token evolution over time.
4. Feed those graph tokens into the original Ctrl-World UNet as `encoder_hidden_states`,
   replacing the original action condition.

The code here is intentionally organized as a separate package so we can iterate without
modifying the original repository.