---
date: 2026-06-16
phase: 3
headline: Figure iteration + website + sim2real plan; cup spill / Keurig deferred to weekends
---

## What we did
- Iterated on **comparison overview figure** (`/data/cameron/scratch_files/para_overview_comparison` — clean side-by-side Traditional Global Regression vs PARA Image-Action Selection, 2 input views, projection step explicit)
- Meeting 10:30–11:30
- Rest of day: new task work + website building + sim2real setup planning

## Reflection
- **Weekend strategy locked**:
  - **Next weekend (Jun 27-28)** post-vacation: grab Keurig station (nobody there weekends) → start long-horizon Keurig task
  - **Cup spill cleanup**: do it next weekend too if it doesn't fit today (Panda team not in on weekends)
  - This is a good use of weekend access — shared stations are "free" only when their primary owners aren't around
- **Sim2real approach sketched today (real-to-sim re-render)**: use known camera intrinsics/extrinsics + known YCB object meshes to re-render a real demo trajectory in sim. Augments the real demo with a sim duplicate. Different from standard sim2real (train sim → deploy real); this is "given a real trajectory, produce its sim image-equivalent." Cheaper to set up because the trajectory is already correct.
  - Useful for: data augmentation (one demo becomes two), domain-robustness eval (does PARA generalize across render styles of the same trajectory?), cheap sim eval infrastructure.
  - Implementation needs: YCB meshes, MuJoCo or Blender renderer, scene alignment to the YAM workspace.
  - Vacation candidate — fits the parked sim2real workstream nicely.
- **Vacation in 2 days**. Pre-trip checklist not yet done — SSH chain verification, durable runs queued, etc.

## Next steps
- Try cup spill cleanup task today if station available; otherwise → next weekend
- Continue website + figure work
- Wed Jun 17: pre-trip checklist (SSH chain from home WiFi, durable training runs queued)
- Vacation Thu Jun 18 → Mon Jun 22: MuJoCo sim2real + website V2 + figure tightenings