# Task B Object Classifier First Pass

This first-pass classifier recognizes Task B RGB crops as `sugar`, `mustard`, `banana`, or `unknown`.

Artifacts:

- Dataset: `outputs/datasets/task_b_object_crops/first_pass_20260624`
- Dataset alias: `outputs/datasets/task_b_object_crops/latest`
- Model: `outputs/models/task_b_object_classifier/first_pass_20260624`
- Model alias: `outputs/models/task_b_object_classifier/latest`
- Checkpoint: `checkpoint.pt`
- TorchScript: `model_ts.pt`
- Mapping: `class_mapping.json`
- Metrics: `metrics.json`
- Visual report: `prediction_contact_sheet.png`

Dataset source:

`scripts/generate_task_b_object_crops.py` launches Task B in a standalone offline scene, renders isolated objects with the legal head camera, and saves RGB crops.  The saved dataset contains 80 sugar, 80 mustard, 80 banana, and 80 background/unknown crops.  The generator uses RGB-D observations to form crop boxes, but the trained classifier input is RGB only.

Legal boundary:

Simulator ground truth is used only offline to write labels from the known Task B object index ranges: object 1-6 sugar, 7-12 mustard, and 13-18 banana.  Future runtime policy code should load the TorchScript model on a legal camera RGB crop and must not read simulator object pose, object id, or labels.

Commands:

```bash
micromamba run -n atec2026 python scripts/generate_task_b_object_crops.py --headless --samples_per_class 80 --background_samples 80 --output outputs/datasets/task_b_object_crops/first_pass_20260624 --seed 24 --settle_steps 5
micromamba run -n atec2026 python scripts/train_task_b_object_classifier.py --data outputs/datasets/task_b_object_crops/first_pass_20260624 --output outputs/models/task_b_object_classifier/first_pass_20260624 --epochs 28 --batch_size 64 --seed 24
```

Validation:

Held-out crop validation accuracy is 96.875% (62/64).  The model has 8,140 parameters and the TorchScript file is about 105 KB.

| true \ pred | sugar | mustard | banana | unknown |
|---|---|---|---|---|
| sugar | 16 | 0 | 0 | 0 |
| mustard | 0 | 15 | 1 | 0 |
| banana | 0 | 1 | 15 | 0 |
| unknown | 0 | 0 | 0 | 16 |

Current heuristic context:

An existing offline trace, `outputs/official_reach_tests/seed7_truefixed_16000_classifier_patch/trace.json`, shows the current color/aspect heuristic repeatedly confusing mustard observations as sugar or unknown.  This is not the same held-out crop validation set, but it is the observed failure mode this model is intended to improve.

Runtime loading sketch:

```python
import json
import torch

model = torch.jit.load("outputs/models/task_b_object_classifier/latest/model_ts.pt", map_location="cpu")
model.eval()
mapping = json.load(open("outputs/models/task_b_object_classifier/latest/class_mapping.json"))

# crop_tensor: [1, 3, 96, 96], RGB float in [0, 1], normalized by mean/std 0.5.
pred_idx = int(model(crop_tensor).argmax(dim=1).item())
pred_class = mapping["idx_to_class"][str(pred_idx)]
```