---
date: 2026-06-12
phase: 3
headline: Towel folding eval (height-off grasps) → arch refactor to many more height bins
---

## What we did
- **Towel folding demo eval**: grasps were in the right (u, v) position but **off in the height dimension** — clean failure mode that points at height resolution as the bottleneck
- **Arch refactor in progress** to enable many more height bins (target 256, up from 32):
  - **Drop the per-voxel / per-point MLP**
  - Each 4D query (time + position, which factors to t + height-bin-along-ray) gets encoded, projected to image features at the projecting pixel in each view
  - **Dot product** of query encoding against per-view image feature → heatmap response
  - **Sum activations** across views (no learned fusion, no view selection — soft gradient through both)
- **AI glasses setup** (today)
- **Visualizations + website work**
- New model training → re-eval tomorrow

## Reflection
- The refactor brings the canonical arch closer to the *old* dino_kv KV-factored philosophy — pure bilinear scoring, no per-voxel learning, just dot product. The multi-view extension is "sum dot products across views," which is the cleanest possible generalization.
- This makes the arch even more "purely binding" in spirit: the only learned operation per voxel is the dot product, no MLP between the projection and the score. Stronger version of the same framing.
- Cheap scaling on height bins falls out for free now — 256 bins is no longer compute-hostile. The towel-folding failure (height-off, position-correct) becomes the empirical motivation for the simplification.
- Light Friday work day in volume but substantive in architecture. The lower energy didn't prevent a real refinement.

## Next steps
- Re-eval new model tomorrow with the 256-height-bin variant — does it nail the towel grasp?
- AI glasses — figure out the use case (data collection? viz? something else?)
- Continue visualizations + website
- Family trip in 8 days — towel folding done-right is still the candidate "cool result"