TWO-LAYER HIERARCHY
CONTEXTXt
SYSTEM 1Latent Policyπlatent
SYSTEM 0Whole-Body TrackerDtracker
ACTIONAt
THREE STRUCTURAL LIMITS
01
INTERFACEInsufficient motion interface
Zt = πlatent(Xt)

Compression can discard task-relevant control information before action is determined.

02
TRAININGSeparate training, coupled inference
θ1 ← Llatentθ0 ← Ltrack

The modules learn from separate objectives, then form one execution chain at deployment.

03
EXECUTIONLow-level action boundary
At ∈ Reach(Dtracker)

A frozen tracker can execute only actions inside its learned control distribution.