Day 105: Stop & Reflect #6 — VLA Architectures & Deployment

Phase VII — VLAs: Architecture to Deployment | Week 15 | 2 hours "You've surveyed 8+ VLA architectures, training recipes, sim-to-real, and deployment. Before the final capstone, consolidate."

Previous: Day 104: Hybrid Architectures Day 2
Next: Day 106: World Models for Robot Control
Week: Week 15 Overview
Phase: Phase VII: VLAs
Curriculum: Full Curriculum

VLA Architecture Map

Fill in this comparison from memory:

Model       Year  Params  Action Rep   VLM Backbone  Key Innovation
─────────── ───── ─────── ──────────── ───────────── ──────────────────
RT-1        2022  ___M    ___________  ___________   ___________________
RT-2        2023  ___B    ___________  ___________   ___________________
Octo        2024  ___M    ___________  ___________   ___________________
OpenVLA     2024  ___B    ___________  ___________   ___________________
π₀          2024  ___B    ___________  ___________   ___________________
π₀.5        2025  ___B    ___________  ___________   ___________________
GR-2        2024  ___B    ___________  ___________   ___________________
GROOT N1    2025  ___B    ___________  ___________   ___________________

Concept Check (30 min)

6 Questions — Answer Without Notes

Q1. Draw the RT-2 pipeline from image input to action output. Label each component.

Q2. Explain why Octo uses blockwise causal attention. What would go wrong with full attention?

Q3. π₀ uses flow matching instead of diffusion. Name 3 practical advantages for robot action generation.

Q4. In coarse-to-fine action prediction, the coarse head uses 64 bins. For $\Delta x \in [-5\text{cm}, 5\text{cm}]$, what's the coarse resolution? Why is refinement needed for sub-millimeter tasks?

Q5. You have a 7B VLA running at 3 Hz. You need 20 Hz for dexterous manipulation. List 3 techniques from Day 104 to close this gap.

Q6. Your deployed VLA achieves 85% success in the lab but 60% in a new kitchen. List 3 potential causes and the corresponding fix for each.

Architecture Selection Flowchart (30 min)

Build a decision tree for choosing the right VLA design:

START: What's your task?
  │
  ├── Simple pick-and-place
  │     └── How much real data?
  │           ├── >10K episodes → RT-1 style (small, fast)
  │           └── <1K episodes → OpenVLA + LoRA fine-tune
  │
  ├── Dexterous manipulation
  │     └── How many DOF?
  │           ├── 7 DOF arm → π₀ (flow matching)
  │           └── >14 DOF (bimanual/hand) → π₀ (long chunk)
  │
  ├── Long-horizon tasks
  │     └── Need language reasoning?
  │           ├── Yes → π₀.5 (plan + act)
  │           └── No → Hierarchical controller
  │
  ├── Multi-robot deployment
  │     └── Same task?
  │           ├── Yes → Octo (multi-embodiment)
  │           └── No → OpenVLA + per-robot LoRA
  │
  └── Novel objects / scenes
        └── Need reasoning about objects?
              ├── Yes → RT-2 / OpenVLA (VLM reasoning)
              └── No → Domain randomization + BC

Synthesis Exercise (30 min)

Design Your Ideal VLA

Given everything you've learned, design a VLA for this scenario:

Scenario: A warehouse robot that must: - Pick diverse products from shelves (novel objects) - Follow natural language instructions from workers - Operate at 15+ Hz control frequency - Work across 3 different robot models - Improve from deployment corrections

Document your design:

Component	Choice	Rationale
VLM backbone
Vision encoder
Action representation
Training data
Transfer approach
Inference optimization
Safety mechanisms
Adaptation strategy

Reflection Prompt (30 min)

Write 300+ words: "What surprised me most about the VLA landscape, and what do I think the next breakthrough will be?"

Consider: - The rapid pace of progress (RT-1 to π₀.5 in 3 years) - The convergence on hybrid designs - The open-source movement (Octo, OpenVLA) - What's still missing (generalization, robustness, speed) - Your prediction for the next 2 years

Weeks 14-15 Review Checklist

[ ] Can explain the architecture of RT-1, RT-2, Octo, OpenVLA, π₀, π₀.5
[ ] Understand tokenized vs continuous vs flow matching action representations
[ ] Know the three-stage VLA training recipe (align → co-fine-tune → specialize)
[ ] Can design a domain randomization strategy for sim-to-real transfer
[ ] Understand teacher-student distillation for deployable policies
[ ] Can choose between VLA architectures for a given deployment scenario
[ ] Know the deployment stack: safety + confidence + recycling + adaptation

Connection to the Thread

Final week begins. Day 106: world models for robot control — can a VLA that predicts the future make better decisions? Days 107-109: deployment deep dive. Days 110-112: the final capstone, where you design, build, and evaluate a complete VLA system. Then: curriculum complete.