Operator fusion, as a key performance optimization technique in the deployment of AI models, significantly improves execution efficiency and has been widely adopted in modern AI compilers. However, for cascaded reduction operations involving multiple loops with inter-loop data dependencies, such as the safe softmax followed by GEMM within attention mechanisms, existing compilers lack effective automated fusion and kernel generation capabilities. Although some works have addressed specific instances through hand-crafted fusion strategies, their solutions are limited in generality and difficult to extend to other similar structures. Given the prevalence of such computational patterns in deep learning models, there remains significant untapped potential in achieving general and automated fusion optimization. In this paper, we present a formal theoretical methodology for analyzing cascaded reductions which can fuse them into a single loop and introduce an incremental computation form. Based on this methodology, we design Red uction Fuser (RedFuser), a framework that automatically identifies supported cascaded reduction patterns and generates optimized fused kernels. Experiments show that RedFuser successfully fuses diverse workloads, achieving up to 2× to 5× speedup over state-of-the-art AI compilers and matching the performance of highly optimized hand-written kernels.

In chemical laboratories, robot operation of instruments often relies on structured auxiliary positioning and teaching methods, which are complicated and laborious. Moreover, the single color, sparse textures, and uneven scales of instruments make the accuracy and robustness of existing visual 6-Dof pose estimation methods difficult to meet the requirement of robot operation. Therefore, we propose a novel pose estimation approach for automatic chemical laboratories using multi-level template matching to assist robots in operating instruments. This approach matches the query image with templates step by step from three levels: template, image, and pixel. During the matching processes from global to local, it achieves prediction of 2D key-point coordinates to accurately estimate instrument pose. At the same time, a template searching method based on the distribution pattern of feature points is proposed to ensure the accuracy of template searching in the real world. The experimental results show that our approach has good robustness and accuracy and meets the requirement of robot operation in automatic chemical laboratories. Note to Practitioners—This paper was motivated by the positioning problem of robot operation for instruments in automatic chemical laboratories. Structured auxiliary positioning methods are complicated and most of the pose estimation works are not suitable for practical applications at present. This paper suggests a novel template searching method and a pose estimation approach using multi-level matching for instruments. We calculate the distances, angles and point number of feature points to construct a new feature searching the best matched template with the input image. Then the predefined key pixels of the template are transformed into the input image from coarse to fine matching to ensure the accuracy and robustness. Due to the instrument CAD model, we can obtain the 2D-3D correspondences and calculate pose of the instrument by the PnP method. The experiment results validate that our method is suitable for practical robot operation in automatic chemical laboratories.

This paper proposes an error state classifier based on the Gaussian Mixture Model (GMM) and a compliance controller for the peg-in-hole assembly task. The concepts of error states and error responses are introduced, and the mapping relationship between them is demonstrated. The GMM-based error state classifier maps the contact forces and torques generated in the assembly process to the probability of different error states. Furthermore, a new compliance controller based on error state probability is proposed to precisely adjust the pose error of the peg and hole. To validate the proposed method, a comparison with Admittance Control (ADC), Variable Compliance Center Control (VCC) and Feature-based Compliance Control (FBCC) is conducted in the task of inserting the vial into the centrifugal hole in the chemical experiment. The results show that the proposed method has a higher insertion success rate and less completion time and can significantly reduce the contact forces and torques during the insertion process.Note to Practitioners—This paper presents a novel error state classifier and a compliance controller for peg-in-hole assembly tasks. Compared to traditional methods, the proposed method can significantly reduce the forces and torques during the assembly process. The error state classifier is obtained by fitting the distribution of contact forces and torques under different error states using the GMM via offline training. During assembly, the classifier calculates the probability of different error states according to the forces and torques, and the proposed compliance controller then uses the error state probability and forces and torques to accurately adjust the pose error, thus ensuring assembly success and quality.