# Figure 6: Point Track Pretraining — Design Notes

## Layout: Training Setup (left) → Data Efficiency Curve (right)

### Left side: Cross-Embodiment Training Setup

Two data sources feeding into the model:

**Source 1: Cross-embodiment pretraining (large)**
- Show the "leader arm" (different embodiment) doing demonstrations
- Label: "×100 episodes" (or whatever the actual number is)
- This is the arm-deleted / different embodiment data
- Only 2D point track supervision (heatmap, no height/gripper/rotation)

**Source 2: Real robot fine-tuning (small)**
- Show the actual robot (LIBERO or SO-100)
- Label: "×10 episodes" (small number, emphasizing data efficiency)
- Full supervision (heatmap + height + gripper + rotation)

Arrow from Source 1 → "Pretrain" → Arrow from Source 2 → "Fine-tune" → Trained PARA model

### Right side: Data Efficiency Chart

X-axis: Number of real robot fine-tuning episodes (1, 2, 5, 10, 50)
Y-axis: Success rate (%)

Lines:
- **PARA (pretrained)** — green, solid — starts high even at 1 episode, climbs to near-max by 10
- **PARA (from scratch)** — green, dashed — lower at small N, catches up at 50
- **ACT / Global Regression (from scratch)** — red, dashed — much lower across all N

The key visual: PARA pretrained reaches high success with very few fine-tuning episodes. The gap between pretrained and from-scratch is largest at small N (1-5 episodes) — that's the data efficiency gain from pretraining.

### Numbers needed from backbones agent
- Success rates for PARA (pretrained + fine-tuned) at N = 1, 2, 5, 10, 50 fine-tuning episodes
- Success rates for PARA (from scratch) at same N values
- Success rates for ACT/global regression (from scratch) at same N values
- Or whatever N values the backbones agent tested

## Check backbones agent status
The arm-deletion experiment was sent to the backbones agent on April 10. Check:
- tmux capture-pane -t backbones
- Look for results in /data/cameron/para/ood_libero/ or the backbones outbox