DYNA-2: add detailed architecture, WAM comparison, and insights sections
- Detailed model structure (mixture-of-transformers, hand-pose pseudo-
actions, flow-matching co-training, mermaid diagram) with the key
structural fact: video prediction is a co-training objective DROPPED at
inference (action head never sees z_t) -> reactive real-time policy.
- New 'How DYNA-2 differs from other WAMs' section: comparison table +
three axes (data purity, world-model-at-inference reactive vs
co-generate, fitted transfer law vs ablation) vs DreamZero/omega-0/
Cosmos-Policy/DreamGen/DreamDojo/EgoScale.
- New 'Key insights' section (39/39 future-pred ablation, human-video
vs teleop, free-at-inference world-modeling, threshold emergence,
authors' own lower-bound/compute caveats).
- Refined DYNA-2 entry in Review-World-Models to note reactive decoupling.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
DYNA-2: cite Dyna's own tech report as primary source; replace 'nothing disclosed' framing with actual figures
Primary source corrected from PR-wire release to the company's technical
writeup (dyna.co/dyna-2), which discloses architecture (mixture-of-
transformers, flow matching), power-law fits with R², ablations, and
per-task results. Rewrote the review + Latest-Papers row accordingly:
real scaling fits (zero-shot robot MSE 0.306*D^-0.0713, R2=0.884),
on-robot 20->28->45->53%, lockbox 0->90%, language 35->67->96%, 90x
distillation. Caveat now: self-published, not peer-reviewed, no weights,
model size undisclosed (was: 'no technical report, unverifiable').
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Fix broken wikilinks in Review-Dyna2 §4 table (escape pipe -> \| for GitHub-wiki table cells)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add DYNA-2 in-depth review (Dyna Robotics World-Action Model launch)
Company announcement (Aug 10, 2026), not a paper: WAM on ~1M h human
egocentric video with no robot data in pre-training, joint next-frame+
next-action, claimed first human-to-robot scaling law smooth over
1k->1M h (~50x EgoScale), 87% vs 46% zero-shot over DYNA-1. Reviewed
with an explicit vendor-claim caveat (no technical paper/benchmark/
weights). Filed under Latest Papers; cross-linked from World-Models,
Human-Video-Transfer, Reviews.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>