Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,91 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-4.0
|
| 3 |
+
library_name: easywam
|
| 4 |
+
base_model: nvidia/Cosmos-Predict2.5-2B
|
| 5 |
+
tags:
|
| 6 |
+
- robotics
|
| 7 |
+
- world-action-model
|
| 8 |
+
- imitation-learning
|
| 9 |
+
- cosmos-predict2.5
|
| 10 |
+
pipeline_tag: robotics
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# EasyWAM-MoT-IDM-Cosmos25
|
| 14 |
+
|
| 15 |
+
This repository hosts released EasyWAM-MoT-IDM checkpoints built on [Cosmos-Predict2.5-2B](https://huggingface.co/nvidia/Cosmos-Predict2.5-2B). EasyWAM-MoT-IDM extends the joint dual-DiT design with a teacher-forced conditional-video branch for action prediction. The checkpoints are trained with the [EasyWAM](https://github.com/OpenMOSS/EasyWAM) codebase.
|
| 16 |
+
|
| 17 |
+
## LIBERO
|
| 18 |
+
|
| 19 |
+
### Results
|
| 20 |
+
|
| 21 |
+
Task success rate (%) under the EasyWAM LIBERO evaluation protocol. All results use `state_position: sequence`; higher is better. Models using the same Cosmos-Predict2.5-2B backbone are shown together for comparison, and 🔥 marks the architecture released in this repository.
|
| 22 |
+
|
| 23 |
+
**Full-Parameter**
|
| 24 |
+
|
| 25 |
+
| Model | Spatial | Object | Goal | LIBERO-10 | Avg. |
|
| 26 |
+
| --- | :---: | :---: | :---: | :---: | :---: |
|
| 27 |
+
| EasyWAM-Unified | 97.8 | 99.4 | 97.0 | 93.0 | 96.8 |
|
| 28 |
+
| EasyWAM-MoT | 98.0 | 98.4 | 98.4 | 95.6 | 97.6 |
|
| 29 |
+
| EasyWAM-MoT-Joint | 98.6 | 99.8 | 98.4 | 96.0 | 98.2 |
|
| 30 |
+
| 🔥 EasyWAM-MoT-IDM | 99.4 | 99.4 | 99.8 | 98.4 | 99.3 |
|
| 31 |
+
| EasyWAM-Hidden | 97.2 | 98.8 | 96.0 | 95.0 | 96.8 |
|
| 32 |
+
|
| 33 |
+
**LIBERO-Plus**
|
| 34 |
+
|
| 35 |
+
| Model | Background | Camera | Language | Layout | Light | Noise | Robot | Avg. |
|
| 36 |
+
| --- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
|
| 37 |
+
| EasyWAM-Unified | 72.7 | 63.7 | 90.1 | 82.6 | 89.2 | 72.1 | 79.2 | 78.2 |
|
| 38 |
+
| EasyWAM-MoT | 60.7 | 75.1 | 92.3 | 82.6 | 92.6 | 81.3 | 58.5 | 77.7 |
|
| 39 |
+
| EasyWAM-MoT-Joint | 62.1 | 60.8 | 96.6 | 84.5 | 94.4 | 69.4 | 76.9 | 77.7 |
|
| 40 |
+
| 🔥 EasyWAM-MoT-IDM | 66.4 | 59.8 | 93.8 | 85.7 | 89.0 | 72.3 | 81.9 | 78.4 |
|
| 41 |
+
| EasyWAM-Hidden | 59.5 | 65.9 | 92.6 | 85.0 | 89.1 | 68.3 | 82.5 | 77.8 |
|
| 42 |
+
|
| 43 |
+
LIBERO-Plus results are reported for full-parameter checkpoints only. See the [complete EasyWAM benchmark table](https://github.com/OpenMOSS/EasyWAM/blob/main/docs/results/result.md) for source results and comparisons across backbones.
|
| 44 |
+
|
| 45 |
+
### Download
|
| 46 |
+
|
| 47 |
+
```bash
|
| 48 |
+
hf download OpenMOSS-Team/EasyWAM-MoT-IDM-Cosmos25 \
|
| 49 |
+
libero_mot_idm_cosmos25.pt \
|
| 50 |
+
libero_dataset_stats.json \
|
| 51 |
+
--local-dir ./checkpoints
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
### Evaluation
|
| 55 |
+
|
| 56 |
+
Install EasyWAM, prepare the Cosmos-Predict2.5-2B dependencies using the [backbone guide](https://github.com/OpenMOSS/EasyWAM/blob/main/docs/instructions/backbone/cosmos25.md), and set up the simulator using the [LIBERO evaluation guide](https://github.com/OpenMOSS/EasyWAM/blob/main/docs/instructions/benchmark/libero.md). Then run:
|
| 57 |
+
|
| 58 |
+
```bash
|
| 59 |
+
# Full-parameter checkpoint
|
| 60 |
+
python experiments/libero/run_libero_manager.py \
|
| 61 |
+
task=libero_easywam_mot_idm_cosmos25 \
|
| 62 |
+
ckpt=./checkpoints/libero_mot_idm_cosmos25.pt \
|
| 63 |
+
EVALUATION.dataset_stats_path=./checkpoints/libero_dataset_stats.json
|
| 64 |
+
```
|
| 65 |
+
|
| 66 |
+
The default protocol evaluates all four LIBERO suites with 50 trials per task. Add `MULTIRUN.num_gpus=<gpu-count>` to distribute evaluation across multiple GPUs.
|
| 67 |
+
|
| 68 |
+
## Available Checkpoints
|
| 69 |
+
|
| 70 |
+
- LIBERO full-parameter checkpoint: `libero_mot_idm_cosmos25.pt`
|
| 71 |
+
- LIBERO normalization statistics: `libero_dataset_stats.json`
|
| 72 |
+
- State-token placement: `sequence`
|
| 73 |
+
- Checkpoint format: EasyWAM PyTorch checkpoint (`.pt`)
|
| 74 |
+
|
| 75 |
+
## Project
|
| 76 |
+
|
| 77 |
+
- Project page: [http://openmoss.ai/EasyWAM/](http://openmoss.ai/EasyWAM/)
|
| 78 |
+
- Code: [https://github.com/OpenMOSS/EasyWAM](https://github.com/OpenMOSS/EasyWAM)
|
| 79 |
+
|
| 80 |
+
## License and Citation
|
| 81 |
+
|
| 82 |
+
EasyWAM code is released under the MIT License. Released checkpoints in this repository use CC BY-NC 4.0 and remain subject to the terms of their base models and training data. If EasyWAM is useful in your research, please cite:
|
| 83 |
+
|
| 84 |
+
```bibtex
|
| 85 |
+
@misc{easywam2026,
|
| 86 |
+
title = {EasyWAM: A Unified and Efficient Framework for Training and Evaluating World Action Models},
|
| 87 |
+
author = {EasyWAM-Team},
|
| 88 |
+
year = {2026},
|
| 89 |
+
url = {https://github.com/OpenMOSS/EasyWAM}
|
| 90 |
+
}
|
| 91 |
+
```
|