num_envs: 1024       # MJX shines with many parallel envs
device: auto         # auto = cuda if available, else cpu
dt: 0.002
substeps: 10
history_length: 10   # RMA-style: 10-step window of (obs, action) pairs