Precision force control technology is central to robotic precision grinding, which presents challenges: optimizing parameters, managing significant contact impacts, and handling uncertain environmental conditions. This article focuses on the robotic precision grinding of blades. To mitigate the pulse impact during initial contact, we devised a post online fastest-tracking differentiator. Furthermore, to enhance force control stability during the grinding, we introduced an adaptive optimal admittance control strategy utilizing the linear quadratic regulator (LQR) to determine optimal control parameters. However, solving the Riccati equation in LQR poses a challenge due to the unknown environment. To tackle this issue, we devised a model-based accelerated learning approach based on the Normalized Advantage Function. This method employs a dual network architecture for policy learning and enhances the reward and acceleration mechanism to improve training speed and convergence. The experiment was performed through robotic precision grinding on aero engine blades. Results indicate that our method enhances force control precision by approximately 210% compared to traditional fixed-parameter grinding methods. Postgrinding, blade roughness improved from Ra0.4 to Ra0.3, and profile deviation was reduced from [0.050, −0.070] to [0.049, −0.030].

Accurate pose estimation of textureless objects from RGB images, without relying on depth information, remains a significant challenge in computer vision. This paper introduces an RGB-based 6D pose estimation method for 6-Degree-of-Freedom robotic grasping. A novel network, named the Regionbased Keypoint Detection Transformer, is proposed, utilizing the concept of set prediction. This network outputs numerous set points, establishing a fuzzy representation of keypoints. Set points are interpreted as keypoints through a voting mechanism for PnP-based 6D pose solving and refinement, guiding robotic grasping actions. Specialized losses, including a set-based coordinates loss and a soft index probability loss for multi-label classification, are designed to facilitate the training of set points. Evaluated on the LineMOD dataset, our method achieved an ADD(-S) accuracy of 92.92 % and a 2D projection accuracy of 98.72 %. Additionally, practical robotic grasping experiments were conducted in real-world environments, encompassing static and dynamic grasping, as well as bin-picking tasks. The picking success rate and completion rate reached 93.02 % and 100.00 % respectively, with the former exceeding the optimal baseline method by 2.11 %. These results validate the accuracy and robustness of our method in cluttered environments.

Traditional integer-order admittance control is widely used in industrial scenarios requiring force control, but integer-order models often struggle to accurately depict fractional-order-controlled objects, leading to precision bottlenecks in the field of precision machining. For robotic precision polishing scenarios, to enhance the stability of the control process, we propose a more physically accurate five-parameter fractional-order admittance control model. To reduce contact impact, we introduce a method combining the rear fastest tracking differential with fractional-order admittance control. The optimal parameter identification for the fractional-order system is completed through Pareto optimality and a time–frequency domain fusion analysis of the control system. We completed the optimal parameter identification in a simulation, which is applied to the robotic precision polishing scenario. This method significantly enhanced the force control precision, reducing the error margin from 15% to 5%.