Because of rehabilitants with various characteristics, rehabilitation robot must perceive the rehabilitant states(strength and position) and then adopt corresponding training mode and control strategy. So how to obtain the state of a rehabilitative limb correctly is very significant for a rehabilitation robot during the training. A new method of dynamics parameter identification of limb based on BP (back propagation) artificial neutral network is presented to offer rehabilitation robot dependable information of limb. The simulation results prove that the method of parameter identification can achieve the state of a rehabilitation limb veraciously and robustly. It can suit different rehabilitants at different stages of rehabilitation even if a spasm happens during training.