nano-world-model

What this skill does

# Nano World Model

> Skill by [ara.so](https://ara.so) — Daily 2026 Skills collection.

Nano World Model is a minimalist, batteries-included repository for training video world models using diffusion-forcing transformers. It supports diverse domains (robot manipulation, gaming, simulation), long-horizon autoregressive rollouts, video-to-3D reconstruction, and MPC-style planning via CEM.

---

## Installation

```bash
git clone https://github.com/simchowitzlabpublic/nano-world-model.git
cd nano-world-model
conda env create -f environment.yml && conda activate nanowm
```

Download the I3D model used for FID/FVD evaluation:

```bash
mkdir -p pretrained_models/i3d && curl -L \
    "https://www.dropbox.com/scl/fi/c5nfs6c422nlpj880jbmh/i3d_torchscript.pt?rlkey=x5xcjsrz0818i4qxyoglp5bb8&dl=1" \
    -o pretrained_models/i3d/i3d_torchscript.pt
```

---

## Environment Variables / Path Configuration

Set these before running any command, or define them in `src/configs/local/paths.yaml`:

```bash
export DATASET_DIR=/path/to/dino_wm_data       # DINO-WM envs (point_maze, pusht, ...)
export CSGO_DATA_DIR=/path/to/csgo             # CSGO HDF5 files
export RT1_DATA_ROOT=/path/to/rt1_fractal      # RT-1 LeRobot mirror (optional)
export RESULTS_DIR=/path/to/results            # checkpoints + logs
```

Or create `src/configs/local/paths.yaml`:

```yaml
# src/configs/local/paths.yaml
dataset_dir: /path/to/dino_wm_data
csgo_data_dir: /path/to/csgo
rt1_data_root: /path/to/rt1_fractal
results_dir: /path/to/results
```

---

## Key CLI Commands

The main entry point is `src/main.py` with [Hydra](https://hydra.cc/) config composition.

### Training

```bash
# DINO-WM PushT, NanoWM-B/2 (default best config: pred-v, additive injection, cosine+ZTSNR)
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2

# DINO-WM Point Maze
python src/main.py experiment=dino_wm_point_maze dataset=dino_wm/point_maze model=nanowm_b2

# CSGO with L/2 model
python src/main.py experiment=csgo dataset=game/csgo model=nanowm_l2_csgo

# RT-1 (fractal)
python src/main.py experiment=rt1 dataset=rt1/rt1 model=nanowm_b2

# Override training steps and batch size inline
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    train.max_steps=50000 train.batch_size=16
```

### Evaluation

```bash
# Evaluate a checkpoint (256 samples, seed=42, 250 DDIM steps)
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    mode=eval checkpoint_path=/path/to/checkpoint.ckpt

# Evaluate with custom sample count
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    mode=eval checkpoint_path=/path/to/checkpoint.ckpt eval.num_samples=512
```

### Long-Horizon Rollout

```bash
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    mode=rollout checkpoint_path=/path/to/checkpoint.ckpt \
    rollout.horizon=50
```

### MPC Planning (CEM)

```bash
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    mode=plan checkpoint_path=/path/to/checkpoint.ckpt \
    planning.method=cem planning.horizon=10 planning.num_samples=128
```

### Video → 3D Point Cloud

```bash
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    mode=video_to_3d checkpoint_path=/path/to/checkpoint.ckpt
```

---

## Pretrained Checkpoints

Load from HuggingFace directly:

```python
from huggingface_hub import hf_hub_download

# Download checkpoint
ckpt_path = hf_hub_download(
    repo_id="knightnemo/nanowm-b2-dino-wm-pusht-100k",
    filename="checkpoint.ckpt"
)
```

Available checkpoints:

| Domain | HF Repo | Steps |
|--------|---------|-------|
| Point Maze | `knightnemo/nanowm-b2-dino-wm-point-maze-30k` | 30k |
| Wall | `knightnemo/nanowm-b2-dino-wm-wall-15k` | 15k |
| Rope | `knightnemo/nanowm-b2-dino-wm-rope-15k` | 15k |
| Granular | `knightnemo/nanowm-b2-dino-wm-granular-15k` | 15k |
| PushT | `knightnemo/nanowm-b2-dino-wm-pusht-100k` | 100k |
| RT-1 | `knightnemo/nanowm-b2-rt1-300k` | 300k |
| CSGO | `knightnemo/nanowm-l2-csgo-100k` | 100k |

Use with CLI:

```bash
python src/main.py experiment=dino_wm_pusht dataset=dino_wm/pusht model=nanowm_b2 \
    mode=eval checkpoint_path=$ckpt_path
```

---

## Configuration System (Hydra)

Config is composed from `src/configs/`. Key axes:

```
src/configs/
├── experiment/          # e.g. dino_wm_pusht, csgo, rt1
├── dataset/             # e.g. dino_wm/pusht, game/csgo, rt1/rt1
├── model/               # e.g. nanowm_b2, nanowm_l2_csgo
├── local/
│   └── paths.yaml       # your local paths (gitignored)
```

Override any config key on the command line:

```bash
# Change prediction target (pred-v vs pred-x0 vs pred-eps)
python src/main.py experiment=rt1 dataset=rt1/rt1 model=nanowm_b2 \
    model.pred_target=pred_x0

# Change action injection strategy
python src/main.py experiment=rt1 dataset=rt1/rt1 model=nanowm_b2 \
    model.action_injection=concat

# Change noise schedule
python src/main.py experiment=rt1 dataset=rt1/rt1 model=nanowm_b2 \
    model.noise_schedule=linear model.zero_terminal_snr=false
```

---

## Code Examples

### Loading a Trained Model Programmatically

```python
import torch
from omegaconf import OmegaConf
from src.models import NanoWM  # adjust import to actual module path

# Load config and checkpoint
cfg = OmegaConf.load("src/configs/model/nanowm_b2.yaml")
model = NanoWM(cfg)

checkpoint = torch.load("/path/to/checkpoint.ckpt", map_location="cpu")
model.load_state_dict(checkpoint["state_dict"])
model.eval()
```

### Custom Dataset with DataSource API

```python
# src/wm_datasets/my_dataset.py
from src.wm_datasets.base import DataSource
import torch

class MyRobotDataSource(DataSource):
    """Custom data source following the DataSource API."""

    def __init__(self, data_root: str, split: str = "train"):
        self.data_root = data_root
        self.split = split
        self._load_index()

    def _load_index(self):
        # Build list of (video_path, action_path) tuples
        ...

    def __len__(self) -> int:
        return len(self.index)

    def __getitem__(self, idx: int) -> dict:
        # Must return dict with keys: "video" (T, C, H, W) and "actions" (T, A)
        video = torch.zeros(16, 3, 64, 64)   # float32, [0, 1]
        actions = torch.zeros(16, 7)          # float32
        return {"video": video, "actions": actions}
```

Register in dataset config:

```yaml
# src/configs/dataset/my_robot.yaml
_target_: src.wm_datasets.my_dataset.MyRobotDataSource
data_root: ${oc.env:DATASET_DIR}/my_robot
split: train
```

### Running Autoregressive Rollout in Python

```python
import torch
from src.models import NanoWM
from src.utils.rollout import autoregressive_rollout

model = NanoWM.load_from_checkpoint("/path/to/checkpoint.ckpt")
model.eval().cuda()

# context_frames: (B, T_ctx, C, H, W), actions: (B, T_rollout, A)
context_frames = torch.randn(1, 4, 3, 64, 64).cuda()
actions = torch.randn(1, 50, 7).cuda()

with torch.no_grad():
    rollout_frames = autoregressive_rollout(
        model=model,
        context_frames=context_frames,
        actions=actions,
        num_ddim_steps=250,
        horizon=50,
    )
# rollout_frames: (B, 50, C, H, W)
```

### MPC / CEM Planning Loop

```python
from src.planning.cem import CEMPlanner

planner = CEMPlanner(
    world_model=model,
    horizon=10,
    num_samples=128,
    num_elites=10,
    num_iterations=5,
    action_dim=7,
)

obs = torch.randn(1, 4, 3, 64, 64).cuda()  # current context
best_actions = planner.plan(obs)             # (horizon, action_dim)
```

---

## Design Choices & Ablation Axes

The repo provides clean ablation across three axes (see `docs/training.md`):

| Axis | Options |
|------|---------|
| **Prediction target** | `pred_v` ✓ (best), `pred_x0`, `pred_eps` |
| **Action injection** | `additive` ✓ (best), `concat`, `cross_attn` |
| **Noise schedule** | `cosine + ZTSNR` ✓ (best), `linear`, `c