Laser cutting technology has been applied in laser demolition and fire rescue due to its advantages of high precision, fast speed, high efficiency, non-contact processing, and adaptability to a wide range of materials. Currently, high-power, ultra-long-distance laser cutting equipment is approaching technological maturity. However, existing high-power, long-distance laser automatic cutting systems primarily rely on point-to-point tracking of moving targets, and automatic cutting systems capable of planning cutting trajectories for complex-shaped objects at long distances remain scarce. Therefore, there is an urgent need for a fast, real-time, and high-precision automatic cutting system.

First, an open-loop control system based on an industrial six-axis robotic arm and a monocular zoom camera was designed. In this hardware control system, the camera and robotic arm adopted an eye-in-hand configuration, while the camera and laser were mounted in a coaxial configuration. The software control system employed Zhang’s calibration method, Tsai-Lenz method, and singular value decomposition-based line fitting method, enabling automatic calibration of the camera’s intrinsic and extrinsic parameters, the pose hand-eye calibration of the robotic arm and camera, and laser beam tool center point calibration. Using the camera’s intrinsic parameters, discrete points of the two-dimensional trajectory line drawn on the camera-captured image were reprojected into a trajectory bundle in the world coordinate system. All pose information was then sent to the robotic arm to perform automatic cutting motion. Following theoretical derivation, an experimental platform was established to verify the system by cutting circular trajectories at a distance of 5 m from the target. After analyzing the causes of experimental deviation, a visual servo closed-loop control system based on a two-axis gimbal and monocular zoom camera was redesigned. In this system, the laser was mounted on the gimbal flange, while the camera was fixed beside the gimbal base in an “eye-to-hand” configuration. The software system employed a kernelized correlation filter fused with color features for laser spot tracking, and a two-dimensional Kalman filtering (KF) was introduced to reduce the effects of atmospheric turbulence, system measurement noise, and system latency. Based on the difference between the pixel coordinates of the discrete target trajectory points drawn on the camera-captured image and the current pixel coordinates of the laser spot, the gimbal inverse kinematics was solved using an online estimation method of the image Jacobian matrix. The calculated angle information was sent to the gimbal controller to perform automatic cutting motion. After theoretical derivation, experimental verification was conducted again by cutting a circular trajectory at a distance of 5 m from the target.

The schematic diagram of the world coordinate system-based trajectory point reprojection rays, laser beam, camera optical center, and industrial six-axis robotic arm base coordinate pose for the open-loop control system obtained through experimental verification was shown in Fig.7, and it was consistent with the pose relationship of the actual hardware platform. The laser automatic cutting process was presented in Fig.8. At a distance of 5 m, the average cutting deviation was 2.2 mm, with a maximum deviation of 4.6 mm. This result validated the successful operation of the open-loop control system based on the industrial six-axis robotic arm and monocular zoom camera. Fig.14 illustrated the laser automatic cutting process using a vision-servo closed-loop control system based on a two-axis gimbal and monocular zoom camera. Fig.15 showed the filtering and prediction effectiveness of the KF. The solid red curve represented the raw measured pixel coordinates obtained through laser spot target tracking using kernelized correlation filtering to fuse color features. This curve exhibited significant fluctuations due to the effects of atmospheric turbulence and inherent measurement noise within the system. The green dashed curve represented the estimated pixel coordinates of the laser spot after KF processing. This curve was smoother compared to the red curve, effectively preventing severe vibrations in the gimbal. The blue dotted curve showed the predicted pixel coordinates of the laser spot for the next time step, obtained through KF prediction. Shifting this curve to the right could closely align with the green curve, mitigating the impact of system latency to some extent. At a distance of 5 m, the average cutting deviation was 0.4 mm, with a maximum deviation of 0.9 mm. This result validated the successful operation of the vision servo closed-loop control system based on a two-axis gimbal and monocular zoom camera. A comparison between Fig.8 and Fig.14 indicated that the open-loop control system produced a smoother cutting trajectory but exhibited significant deviation, while the vision servo closed-loop control system showed jitter in the cutting trajectory but with smaller deviation.

Compared with the open-loop control system based on an industrial six-axis robotic arm and a monocular zoom camera, the vision-based servo closed-loop control system based on a two-axis gimbal and monocular zoom camera reduced cutting trajectory deviation by 81.8%. Both approaches have its own advantages and disadvantages. The former requires calibration at different zoom levels and exhibits lower accuracy at long distances, but offers better stability and robustness. The latter eliminates calibration, achieves higher precision, and requires fewer degrees of freedom from the robot, but it is more susceptible to environmental disturbance. These findings provide theoretical guidance for advancing the application of high-power, long-distance lasers in laser demolition and fire rescue operations.