Writeup, demo, share, fresh-clone test

This is a valid v1.0 placeholder page for the later curriculum arc. Full interactive lab treatment ships after Week 1 dogfooding.

LECTURE & READING

Hour 1 — 2-page writeup (45 min)

docs/writeup.md:

# <Track X>: <one-line title>

## Hypothesis
<copy from README>

## Setup
- Robot/env: <Spot Joystick Flat Terrain | LIBERO-Spatial | etc.>
- Algorithm: <PPO + DR + RMA | π0.7 + LoRA r=128 | etc.>
- Data: <50h auto-collected | 30 episodes hand-teleop | etc.>
- Compute: <1× H100, 12 hours total>

## Result (headline)
| Variant | Success rate | n seeds |
|---|---|---|
| Baseline | 0.62 ± 0.07 | 3 |
| Extension | 0.81 ± 0.04 | 3 |

[Plot: headline.png]

The extension produces a 19-percentage-point improvement (p < 0.05 by t-test).

## Ablation
[Plot: ablation.png]

Rank ∈ {16, 64, 128} produced success rates {0.71, 0.81, 0.79}; r=64 is sufficient.

## Failure analysis
[Plot: failure.png]

Most baseline failures (62%) were "grasp slip" — the policy released the object
mid-trajectory. Extension reduced these to 31%, suggesting the increased capacity
captures grasp-stability features the baseline missed.

## Connection to JEPA-family research
[1-2 paragraphs connecting your work to the broader JEPA / world-model debate.]

## What I'd do with 5× compute
- More seeds (5 → 10) for tighter error bars.
- Larger model (300M → 1B for Track B; rank 128 → 512 for Track A).
- Real hardware deployment for Track C/D.

## Reproduce
\`\`\`
git clone <url>
cd <repo>
make install && make data && make train && make eval
\`\`\`
Wall-clock: ~6 GPU-hours, 1× H100. Commit SHA: <abc1234>.

Hour 2 — 8-slide deck + 2-min demo video (75 min)

Slide deck

1. Title + 1-line hypothesis 2. Why this matters (research context, JEPA / / angle) 3. Method (block diagram) 4. Data (samples, scale) 5. Headline result (the bar plot) 6. (the second plot) 7. Failure analysis (the third plot, with takeaway) 8. Reflection (what didn't work, what's next)

Use the Anthropic Imagine style for slides if you have access; otherwise plain Slides / Keynote / reveal.js.

Demo video

A 2-minute screen recording showing:
0:00–0:20 — title + hypothesis card
0:20–1:00 — eval video clips (2-3 successes from extension; 1 failure from for contrast)
1:00–1:30 — show the headline plot
1:30–2:00 — show make reproduce running successfully

Save as videos/demo.mp4.

LAB

Hour 3 — Fresh-clone test + share (60 min)

Fresh-clone

cd /tmp && rm -rf cap-test
git clone <your-w7-capstone-url> cap-test
cd cap-test
make install
make eval  # uses your saved checkpoints

Headline number must match your writeup's reported number within ±1%.

If it doesn't, fix it before publishing. This is rubric category 3.

Full source continues in the committed curriculum files. The v1.0 page exposes the day flow and lab surface without inventing content.

Completion controls unlock when this day graduates from placeholder to full lab.

Optional next build

Keep going with a bonus project

Start the HUG + YouTube task robot project when you want a hardware-heavy follow-up: video-to-task planning, RGB-D grasping, retargeting, and honest robot evaluation.

Open bonus projects