TRAIN · SEPARATE UPDATES
SYSTEM 1
Latent Policy
π
latent
(θ
1
)
UPDATE
θ
1
← L
latent
NO SHARED
GRADIENT
SYSTEM 0
Whole-Body Tracker
D
tracker
(θ
0
)
UPDATE
θ
0
← L
track
INFER · ONE SERIAL PATH
CONTEXT
X
t
SYSTEM 1
π
latent
Zₜ
SYSTEM 0
D
tracker
ACTION
A
t