In the case of some elderly whose motor ability is deteriorated, the environment perception ability is sometimes deteriorated also. In order to deal with that problem, the power-assist robot with the perception-assist, which assists not only the user's motion but also the user's interaction with an environment, has been proposed. The power-assist robot with the perception-assist has some sensors in order to monitor the interaction between the user and the environment. The robot tries to modify the user's motion automatically during the power-assist if the robot recognizes any problems in the user's motion. A human has a dexterous hand with many DOFs, and performs tasks while using many tools in the hand. Since many kinds of tools are used in daily activities, the undesired motion such as a risky behavior to the person is different between the tasks or tools. From these reasons, when the power-assist robot carries out the perception-assist, the robot needs to perform the perception-assist based on the user's motion and the grasped tool by the user. However, it is not realistic to prepare the perception-assist required for every tool or task because there are many kinds of tools and tasks. In this paper, the perception-assist for an upper-limb power-assist robot is discussed. The robot learns the properly required perception-assist based on EMG signals. The effectiveness of the proposed method has been evaluated by the performing experiments.