README.md

# Sim2Sim2Sim Checkpoints

Official pretrained checkpoints for [Dynamics Distillation for Efficient and Transferable Control Learning](https://arxiv.org/abs/2605.01516).

## Directory Structure

```
dynamics_model/              # Stage 1: Learned dynamics models
├── beamng_bicycle_*/        # Bicycle model variants
├── beamng_ddm_*/            # Deep Dynamics Model (DDM) variants
├── beamng_trans_*/          # Transformer-based models
├── beamng_dytr_*/           # DYTR (Residual Learning) models
└── beamng_manual_PID_*/     # Manual PID control baselines

control_policies/            # Stage 2 & 3: Trained RL policies
├── PPO____*_bicycle/        # Policies trained on bicycle dynamics
├── PPO____*_ddm/            # Policies trained on DDM dynamics
├── PPO____*_trans/          # Policies trained on Transformer dynamics
├── PPO____*_dytr_ddm/       # Policies trained on DYTR dynamics
└── PPO____*_oracle/         # Oracle policies (full-state observation)
```

## Usage

### Download Specific Model

```python
from huggingface_hub import hf_hub_download

# Download a dynamics model
dynamics_model = hf_hub_download(
    repo_id="alfredgu001324/Sim2Sim2Sim",
    filename="dynamics_model/beamng_trans_10/best_model.pt"
)

# Download a trained policy
policy = hf_hub_download(
    repo_id="alfredgu001324/Sim2Sim2Sim",
    filename="control_policies/PPO____R_80000__11_25_10_41_44_840_trans/model_PPO____R_80000__11_25_10_41_44_840_001280.pt"
)
```

### Batch Download

```python
from huggingface_hub import snapshot_download

# Download all checkpoints
local_dir = snapshot_download(
    repo_id="alfredgu001324/Sim2Sim2Sim",
    cache_dir="./ckpts",
    repo_type="model"
)
```

## Model Details

### Dynamics Models (Stage 1)

- **Bicycle**: Simple analytical model serving as baseline
- **DDM**: Deep Dynamics Model - neural network trained on BeamNG data
- **Transformer**: Sequence-aware dynamics model using transformer architecture
- **DYTR**: Residual Learning towards High-fidelity Vehicle Dynamics Modeling with Transformer

### Control Policies (Stage 2 & 3)

All policies trained using PPO with 80,000 environment steps:
- **trans**: Trained on Transformer dynamics model
- **ddm**: Trained on DDM dynamics model  
- **dytr_ddm**: Trained on DYTR-wrapped DDM dynamics model
- **bicycle**: Trained on bicycle model
- **oracle**: Full-state observation policies
- **FF**: Feed-forward policies
- **trans_cond**: Transformer with condition encoding for surface changes

## Citation

If you use these checkpoints, please cite the paper:

```bibtex
@article{GuChittaEtAl2026,
  author    = {Gu, Xunjiang and Chitta, Kashyap and Golchoubian, Mahsa and Suplin, Vladimir and Gilitschenski, Igor},
  title     = {Dynamics Distillation for Efficient and Transferable Control Learning},
  journal   = {arXiv preprint arXiv:2605.01516},
  year      = {2026}
}
```

## License

Apache 2.0