a) Teaching function. Precisely teaching for complicated weld seam of work- pieces is a tedious work either online manipulating or offline programming. It is con- venient to teach few points grossly to obtain approximate track of seam for automatic manufacturing in weld industry. Therefore, the grossly teaching mode is adopted for the welding robot to sample weld seam. For example, it is sufficient to teach only two or three points; for complicated weld seam for simple beeline seam; for girth weld seam of pipe, the parameters of the girth seam (e.g. circle or ellipse), or the bounda- ries are needed.

b) Multi-axis linkage. As the external axis of welding robot, some auxiliary mechanism such as rollers or positioning machines will be controlled for multi-axis linkage. The motion and velocity of the external axis are sampled and controlled to estimate the position and pose of workpieces, and the amount of feed of weld torch can be calculated with relative position and pose between weld seam and the weld torch.

c) Motion control. Coordinated motion control of the macro and micro mecha- nisms is realized by intelligent algorithms in this function, which will be discussed in the next section.

d) Monitoring and human-compute interface. A stereo cameras and structured light vision sensor are mounted on the end-effector of the robot. The position of weld seam can be visually measured by the structured light vision sensor for seam tracking in process of weld; the relative pose of wed torch and workpieces is inspected by the stereo cameras for the alignment of the torch. The images of workplace are shown on the console in real time, and the results of image process and features extraction are shown to monitor the status of weld. The welding parameters and the welding devices can be managed in the function module.

e) Manual manipulation. The pose and position of the weld torch can be adjusted by manual control with 6 translational joints and 3 rotational ones. By manual func- tion, the torch and vision senor can reach desirable pose, and the torch are moved near the initial weld position. Manual function is also needed when debugging and mainte- nance of devices.

3.2 Control Architecture

Control system of welding robot is designed with multi layers, which include the interactive and intelligent layer, motion planning layer, motion control and servo control layer. The Control architecture of welding robot is shown in Figure 2. Con- cretely, an industrial PC is employed for the host computer; two motion control cards plan the multi-axis linkage; servo drivers and servo motors with the rotary encoders comprise driving control units. The functions of top two layers are implemented by the host computer, and bottom two layers are implemented by the motion control cards, the servo drivers and the motors.

Fig. 2. Architecture of Control System

The top layer is interactive and intelligent layer, which provides the management functions, including human-machine interaction, task planning, control of welding system, the signal processing of field sensors and vision sensors, monitoring and intelligent control. In the layer, the welding system is started or closed via the com- munication interface of the host computer, teaching task is accomplished, and the control variables of weld torch are calculated by intelligent algorithms.

The second layer is motion planning. According to the control variables or the de- sirable pose of weld torch given by top layer, the motion of the weld torch are planned in Cartesian space, and the variables of the joints are calculated in joint space by in- verse kinematics of the welding robot.

Given variables of the joints in motion planning layer, the analog signal output is calculated in the motion control layer, with the feed back of the encoder of the mo- tors. The close loop control is implemented by the motion control cards for multi-axis linkage. The commands of motion control are buffed in motion planning layer, and the synchronization and coordination of translational joints and rotational ones is also realized in the layer.