num_envs: 64
device: auto  # auto = cuda if available, else cpu
dt: 0.002
substeps: 20
action_ema_alpha: 0.2  # motor smoothing (τ ≈ 179ms, ~5 steps to reverse)