# Mult-skill ACT Models

Multi-skill Action Chunking Transformer models for Piper dual-arm robot.

## Models

| Model | Description |
|-------|-------------|
| `point_classifier/` | ResNet-18 point classifier (10 classes: point1-point10) |
| `exp1_all_points/` | ACT policy trained on all 150 episodes (10 points combined) |
| `exp2_point1/` ~ `exp2_point10/` | ACT policies trained per-point (15 episodes each) |
| `exp3_act_x/` | Experimental ACT policy |

## Normalization

`normalization/` contains pre-computed normalization constants for each dataset:
- `all_points_v30_norm.py` - Stats from combined dataset
- `point1_v30_norm.py` ~ `point10_v30_norm.py` - Stats from per-point datasets

## Usage

### Download all models
```bash
pip install huggingface_hub
huggingface-cli download QRP123/mult-skill-act-models --repo-type model --local-dir ./models
```

### Point Classifier Inference
```python
from point_classifier_model import PointClassifier
import torch

checkpoint = torch.load("point_classifier/point_classifier_best.pth")
model = PointClassifier(num_classes=10, pretrained=False)
model.load_state_dict(checkpoint["model_state_dict"])
```

## Robot Setup
- Robot: Piper dual-arm (7D + 7D = 14D state/action)
- Cameras: fisheye_left, fisheye_right, wide_mid, wide_top (256x256)
- FPS: 30