---
license: other
license_name: nvidia-open-model-license
license_link: https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/
library_name: pytorch
tags:
  - latent-action-model
  - world-model
  - robotics
  - video-prediction
  - research
---

# CD-LAM

This repository contains three CD-LAM model entries.

| entry | file | role |
|---|---|---|
| LAM | `models/lam/model.pt` | 32D latent-action model |
| Pretrain | `models/pretrain/model.pt` | latent-action-conditioned world-model overlay |
| Posttrain | `models/posttrain/model.pt` | robot-action post-trained world-model overlay |

The posttrain folder also contains its required
`models/posttrain/bridge.pt` and
`models/posttrain/action_contract.json`. They are support files for the
posttrain entry, not additional models.

## Use with the source release

```bash
git clone https://github.com/yufanwei/CD-LAM.git
cd CD-LAM
bash setup.sh --accept-base-license --with-models
```

The source downloader pins an immutable Hub revision and verifies file sizes
and SHA-256 values from `asset_manifest.json`.

## Compatibility

The LAM and pretrain entries are compatible with each other. The posttrain
entry must use the bridge and action contract in the same folder.

The two world-model files are overlays and require the compatible external
NVIDIA 2B base checkpoint, Cosmos tokenizer, and text encoder. Exact revisions
are recorded in `asset_manifest.json` and the
[CD-LAM source release](https://github.com/yufanwei/CD-LAM).

## Scope

These are inference and evaluation research artifacts, not optimizer-resume
checkpoints, a robot policy, or a safety controller. Model identity and tensor
verification do not by themselves reproduce the manuscript tables.

## License and citation

Read `NOTICE` and `LICENSE.pdf` before using the checkpoints. Built on NVIDIA Cosmos.
The paper BibTeX entry is available in the source repository's
[README](https://github.com/yufanwei/CD-LAM#citation).