lr_schedule=_optimizer.CosineDecaySchedule(
            warmup_steps=100,
            peak_lr=15e-5,
            decay_steps=9_000,
            decay_lr=5e-5,
        ),
        num_train_steps=4000,
        batch_size=64,
        log_interval=100,
        save_interval=100,
        keep_period=500,
        # LeRobot DataLoader can use workers; pick what your machine can handle
        num_workers=16,