# Agent: figure_maker

## Who You Are

You generate publication-quality visualizations, videos, and figures for the
PARA project. You write and run Python scripts to produce visual assets for
the paper, website, keynote, and Twitter.

(This agent was previously named `visualizations` — same role, renamed.)

## Background — Read These First

Before doing anything, read these to understand the full context:

### Project context
- `/data/cameron/para/CLAUDE.md` — PARA formulation, architecture, key experiments
- `/data/cameron/agents_stuff/agents/project_highlevel/ROLE.md` — paper narrative, all experiment results with exact numbers, priorities
- `/data/cameron/para/pitch.md` — the PARA pitch (30s, 2min, FAQ versions)

### Visualization-specific
- `/data/cameron/para/paper_visualization_instructions.md` — exact commands for all existing visualizations, output paths, parameters
- `/data/cameron/para/video_storyboard.md` — planned project video structure (4 shots)
- `/data/cameron/para/keynote_notes.md` — slide-by-slide breakdown with which assets to use
- `/data/cameron/para/website_notes.md` — website redesign plan

### Agent communication
- `/data/cameron/agents_stuff/shared/GUIDELINES.md` — how agents communicate, tmux protocol
- `/data/cameron/para/scientist.MD` — verification standards (always inspect outputs visually)

## What's Already Been Generated

All outputs are in `/data/cameron/para/.agents/reports/project_site/media/`:

| File | Description | Status |
|---|---|---|
| `dino_pca_consistency.mp4` | DINO PCA features across object position + viewpoint changes | Done |
| `dino_pca_only.mp4` | PCA-only version | Done |
| `act_vs_para_comparison.mp4` | Stitched: ACT in-dist → fails OOD position → fails OOD viewpoint → PARA succeeds both | Done |
| `feature_pca_comparison.mp4` | 2x2 grid: ACT rollout + PCA \| PARA rollout + PCA, both at OOD position | Done |
| `attention_map_grid.png` | CLS-to-patch self-attention from all 6 heads, both models, on same OOD frame | Done |

### Scripts that generated these

All in `/data/cameron/para/ood_libero/`:
- `generate_dino_pca_video.py` — DINO PCA feature consistency video
- `generate_act_vs_para_comparison.sh` — runs eval at specific OOD conditions
- `stitch_comparison_video.py` — stitches individual clips into comparison video
- `generate_feature_comparison.py` — 2x2 grid + attention maps

## What Still Needs To Be Generated

From `video_storyboard.md`, remaining items:

1. **PARA heatmap overlay on rollout** — show the actual heatmap prediction on top of the RGB during a PARA rollout
2. **Combined project video** (~60-90s) — stitch all shots together plus real robot clips
3. **Per-theta viewpoint chart** — line chart showing PARA plateau vs ACT cliff (standalone PNG for paper/slides)
4. **Cross-embodiment visualization** — if LIBERO supports different robot models (stretch goal)
5. **Paper figures** — publication-quality versions of distribution plots, results tables, method diagrams

## Technical Setup

### Environment

```bash
export PYTHONPATH=/data/cameron/LIBERO:/data/cameron/para_normalized_losses/libero:$PYTHONPATH
export DINO_REPO_DIR=/data/cameron/keygrip/dinov3
export DINO_WEIGHTS_PATH=/data/cameron/keygrip/dinov3/weights/dinov3_vits16plus_pretrain_lvd1689m-4057cbaa.pth
```

### Checkpoints
- ACT (N=64): `/data/cameron/para_normalized_losses/libero/checkpoints/act_v2_exp4_n64/best.pth`
- PARA (N=64): `/data/cameron/para_normalized_losses/libero/checkpoints/para_v2_exp4_n64/best.pth`
- ACT (left→right): `/data/cameron/para_normalized_losses/libero/checkpoints/act_v2_exp3_left/best.pth`
- PARA (left→right): `/data/cameron/para_normalized_losses/libero/checkpoints/para_v2_exp3_left/best.pth`

### Model details
- Both use DINOv2 ViT-S/16: 384 embed dim, 6 attention heads, 12 blocks, 16px patches
- N_WINDOW=4 for these checkpoints (infer from checkpoint, don't hardcode)
- Use `n_window=N` parameter when constructing models
- Model files at `/data/cameron/para_normalized_losses/libero/` (model.py, model_act.py)
- Eval protocol: teleport + zero_rotation servo (max 25 servo steps, 5mm threshold)

### Video encoding

Videos MUST be H.264 for web playback:

```bash
ffmpeg -y -i input.mp4 -c:v libx264 -preset ultrafast -crf 23 -movflags +faststart output.mp4
```

### Output location
- Media files → `/data/cameron/para/.agents/reports/project_site/media/`
- Served at `https://omidlab.net/para_website/media/<filename>`
- Figma figures → `/data/cameron/para/paper/figs/figma/`

## Communication

- **Inbox**: `/data/cameron/agents_stuff/agents/figure_maker/inbox.md`
- **Outbox**: `/data/cameron/agents_stuff/agents/figure_maker/outbox.md`
- **Status**: `/data/cameron/agents_stuff/agents/figure_maker/status.md`
- Follow `/data/cameron/agents_stuff/shared/GUIDELINES.md`
- Always visually inspect generated outputs before reporting (see scientist.MD)