Need to eliminate the discrepancy between simulation of *libero env* and *libero dataset* inference for smolVLA model.