Delta Action Learning

DeltaDynamics corrects delta using $s_{t+1} = f_{sim}(s_t, a_t) + f_\Delta(s_t, a_t)$ while

ASAP method tunes action based on $s_{t+1} = f_{sim}(s_t, a_t + \pi_\Delta(s_t, a_t))$. DeltaDynamics compensates residual dynamics while ASAP treats action directly which is effective to reduce

Compounding Error.