The force-length relation of muscles is one of the basic mechanical properties. However, there have been few reports about the relationship between joint angle and torque for a human upper extremity. The purpose of this study is to determine the relationship between joint angle and isometric torque of human upper extremities. A nonlinear model (neural network model) is used to estimate joint torque under the restriction of constant muscle activities. Three normal volunteers were seated in a chair and put their upper extremities on hanging plates which were allowed to move freely on the horizontal plane. External force was applied to the wrist with a string, tied to their wrist, with a weight. External force was 0, 2.5, 5, 7.5, 10, 12.5 [N]. Subjects were instructed to maintain their upper extremity at a target joint angle and to balance against the external load with minimum effort. Target shoulder joint angle was 30°, and elbow joint angle was 30°, 35°, …, 100°. Five-channel integrated EMG (IEMG), joint angles, and torque were recorded for 5 sec. Elbow and shoulder joint angles were calculated by detecting positions of infrared LEDs attached on the upper extremity of a subject with a 3D motion analyzer (OPTOTRAK, Northern Digital Inc.). Surface Ag-AgCl electrodes taped to the skin were used for bipolar EMG recording. Five-channel EMGs were recorded from caput longum muscle biceps brachii, caput laterale, muscle triceps brachii, muscle pectoralis major, and muscle deltoideus. IEMG was obtained with a differential amplifier, a rectifier circuit, and a second-order low pass filter. External force was measured with a strain gage. Elbow and shoulder joint torque were calculated geometrically from joint angles, and arm length, and the external force. The neural network model consists of an input layer of 7 units, one hidden layer of 6 to 15 units, and an output layer of 2 units. The inputs were 5ch IEMGs and both elbow and shoulder joint angles. The outputs were elbow and shoulder joint torque. After learning, elbow joint torque was estimated where 5ch IEMGs were fixed and only elbow joint angle was varied. Estimated results were as follows. Torque of the elbow extensors monotonically increased as flexing of elbow joint. However, the torqueangle relation of flexors had a peak at about 90°. These relationships were discussed by considering the dependence of muscle force on the muscle length, as well as the angle dependence of moment arms and muscle length of the elbow flexor and extensor muscles.