README.md

---
license: other
license_name: openmdw-1.1
license_link: LICENSE
pipeline_tag: robotics
base_model: nvidia/Cosmos3-Edge
tags:
- robotics
- world-model
- world-action-model
- cosmos
---

# GreenWAM: A World Foundation Model for Robotics

[Paper](https://arxiv.org/abs/2610.18774) · [Project page](https://greenvla.github.io/GreenWAM/) · [Code](https://github.com/greenvla/GreenWAM)

A robot needs a world model that predicts how its actions change the scene and a policy that chooses those actions.
**GreenWAM** is one robot world model that serves both roles across different bodies. It is built on Cosmos 3 Edge
and trained with a three-part recipe:

- **Shared anatomical action space.** 20 robot families, from bimanual arms to dexterous humanoids, are described
  in common joint coordinates: every native joint is mapped to its anatomical role and to one of 79 shared slots.
- **Joint robotics training.** One model learns these families in visual, action-conditioned and world–action modes.
- **Action expert.** Policies split an action expert off the generator; it shares attention with it in every layer.

This repository holds the four benchmark checkpoints of the [technical report](https://arxiv.org/abs/2610.18774). The code that evaluates them (policy
servers, benchmark clients, the WorldArena rollout, evaluator and scoring) is in the
[GreenWAM repository](https://github.com/greenvla/GreenWAM).

## Checkpoints

| Folder | Benchmark | Role | Reported result |
|---|---|---|---|
| [`robotwin`](./robotwin) | RoboTwin 2.0 (50 tasks) | policy | **67.54** mean · 86.98 clean · 48.10 randomized |
| [`libero`](./libero) | LIBERO · LIBERO-plus | policy | **97.4** LIBERO · **86.8** LIBERO-plus zero-shot |
| [`calvin`](./calvin) | CALVIN ABC→D | policy | **4.64** average length |
| [`worldarena`](./worldarena) | WorldArena 1.0, track 1 | world model | **63.31** EWMScore |

Every folder holds the EMA weights (bf16 safetensors), the `config.yaml` the evaluation reads, the action statistics
of its benchmark and a `SHA256SUMS` file.

## Usage

Download one checkpoint:

```bash
hf download SberRoboticsCenter/GreenWAM --include "robotwin/*" --local-dir checkpoints/GreenWAM
```

The checkpoints also need the Wan2.2 VAE and the Cosmos 3 Edge text tokenizer at the revision they were evaluated
with:

```bash
hf download Wan-AI/Wan2.2-TI2V-5B Wan2.2_VAE.pth --local-dir checkpoints/Wan2.2-TI2V-5B
hf download nvidia/Cosmos3-Edge --revision a9d944e2c6a1bf9f48b92ad16348e70c5f1836ba --local-dir checkpoints/Cosmos3-Edge
export WAN_VAE_PATH=$PWD/checkpoints/Wan2.2-TI2V-5B/Wan2.2_VAE.pth COSMOS3_EDGE_DIR=$PWD/checkpoints/Cosmos3-Edge
```

Installation, the policy servers and the evaluation of each benchmark are described in the
[GreenWAM repository](https://github.com/greenvla/GreenWAM).

## License

The weights are released under the [OpenMDW License 1.1](./LICENSE), the license of Cosmos 3 Edge.

## Citation

```bibtex
@misc{sidorov2026greenwamworldfoundationmodel,
  title         = {GreenWAM: A World Foundation Model for Robotics},
  author        = {Gennady Sidorov and Valerii Kornilov and Denis Grankin and Polina Fedotova and Aleksey Postnikov},
  year          = {2026},
  eprint        = {2610.18774},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.18774}
}
```