---
date: 2026-06-15
phase: 3
headline: Backbone ablation expanding — DynaFlip beats DepthAnything; PaliGemma trails
---

## What we did
- **Two more backbones trained + evaluated**:
  - **PaliGemma** (VLM backbone) — works fine, but **worse than both image-based backbones**
  - **DynaFlip** (robotics-specific backbone) — **beats DepthAnything**
- The Phase 3 backbone ablation table is forming organically (originally planned for backbones-agent on lab GPUs; happening at TRI instead)
- Egg + teapot models from Sun continue iterating

## Reflection
- DynaFlip > DepthAnything is a real result — extends the "head > backbone" story onto a *different* family of backbones. The contribution isn't just "DINO+head beats DINO-L" anymore; it's becoming "the head works across backbone families and the right one depends on task."
- PaliGemma underperforming image-only backbones is interesting and worth noting carefully: VLM features have language structure that may not be the right inductive bias for action prediction. Don't pitch this as "PaliGemma is bad" — pitch as "language-aligned features don't dominate where pixel-aligned features do." Both are useful framings, different audiences.
- Backbone ablation suite as of today: DINOv3-S/16+ (canonical) · DepthAnything · DynaFlip · PaliGemma. Solid range for a table.

## Next steps
- Continue iterating egg + teapot
- Try cup spill cleanup task tomorrow if station is free (otherwise next weekend)
- Comparison overview figure iteration (good draft at `/data/cameron/scratch_files/para_overview_comparison`)