Perpetual Humanoid Control for Real-time Simulated Avatars

They talk about how learn all the primitives at the same time results in forgetting.

Uses many things from AMP.

First, we learn a set of primitives $p^k$. Then, we use a high level controller to switch between them

• How is the composer implemented?

“Our composer $C(w_t^{1:K+1}|s_t)$ consumes the same input as the primitives and outputs a weight vector w1:K+1 t ∈ Rk+1 to activate the primitives.”