metadata
title: Visuo-Tactile Stage-1 Gate (Cosmos)
emoji: 🧪
colorFrom: indigo
colorTo: gray
sdk: static
pinned: false
Stage-1 gate: does the action-motion encoder transfer to Cosmos latents?
Verdict: NO TRANSFER — the encoder learns well (D beats zero by 32.3%, C by 33.0%), but the action map adds NOTHING over tactile-only (D is 1.0% worse than C; Wan2.2 had D +3.3% over C). The conditioning signal Stages 2-3 inject is redundant here.
Active-pixel EPE (lower is better) for the action-motion encoder trained on Cosmos cv4x8x8 tactile latents, against the zero-motion baseline on the same val split and against the published Wan2.2 numbers.
See data.json for the raw metrics.