Goal-conditioned dual-action imitation learning for dexterous dual-arm robot manipulation