From geometric analysis to semantic reasoning: the evolution of robotic grasping perception paradigms

Robotic grasping perception serves as a fundamental prerequisite for autonomous manipulation and embodied intelligence, acting as a core technology for robots to interact with the physical world. As application scenarios expand from structured industrial assembly lines to unstructured environments such as households and logistics, the challenges facing grasping tasks have become increasingly complex. Modern robots are required not only to perceive the geometric attributes of objects in scenes characterized by clutter, occlusion, and varying lighting but also to understand semantic information and task-specific contextual constraints. Despite the proliferation of research, existing literature often focuses on single technical branches or physical stability assessments, lacking a systematic tracing of the evolutionary logic of perception paradigms. Bridging this gap is essential for developing general-purpose robotic systems that possess high generalization performance and robust task comprehension.

The evolution of robotic grasping perception can be categorized into three logically progressive stages: analytical geometry-driven methods, visual data-driven methods, and semantic understanding and reasoning enhancement. The analytical geometry-driven paradigm represents the classical approach, focusing on deriving optimal grasp configurations through explicit mathematical definitions and geometric features. This paradigm includes three main categories. First, stability analysis based on contact mechanics models grasping as a set of contact points on an object's surface, using criteria like force-closure and form-closure to resist external disturbances. Second, explicit computation based on geometric features utilizes local geometric information, such as polyhedral models and grasp planes, to infer configurations without requiring complete physical models. Third, task-oriented grasping relies on predefined rules and tool affordances, where robots learn the effects of different actions on objects to complete tasks. While providing a theoretical foundation, these methods are often limited by their sensitivity to sensing noise and their dependency on precise physical parameters. The visual data-driven paradigm has emerged as the mainstream, shifting away from explicit modeling toward learning complex strategies from massive "experience" data. This paradigm leverages deep neural networks to establish direct mappings from visual observations to grasp poses. Key methodologies include imitation learning, which learns from expert demonstrations or template matching; two-stage paradigms that decouple grasp synthesis into candidate generation and quality evaluation; and end-to-end direct prediction methods. Recent advancements have also explored the use of 3D representations like Signed Distance Fields (SDF), Neural Radiance Fields (NeRF), and 3D Gaussian Splatting to capture high-order geometric features. The frontier of the field is currently defined by the semantic understanding and reasoning enhancement paradigm. Driven by the development of Large Language Models (LLMs) and Vision-Language Models (VLMs), this paradigm seeks to move beyond physical stability to achieve deep understanding of the physical world. It includes multimodal fusion approaches that align natural language instructions with visual features, hierarchical planning guided by high-level semantic reasoning, and end-to-end foundation models. For instance, Vision-Language-Action (VLA) models like Grasp VLA utilize billion-scale synthetic data to achieve zero-shot or few-shot generalization across diverse objects and tasks. In tandem with these paradigms, grasping datasets have evolved from early planar benchmarks like the Cornell Grasping Dataset to large-scale 6-DoF real-world benchmarks like GraspNet-1Billion and massive synthetic datasets like SynGrasp-1B. Evaluation systems have also matured, utilizing metrics such as success rate for task reliability and average precision for proposal accuracy.

Robotic grasping perception undergoes a profound transformation from rigid geometric analysis to intelligent, semantic-driven reasoning. This evolution significantly enhances the generalization capabilities and task adaptability of robots in complex, unstructured environments. However, several common technical bottlenecks remain. Bridging the "Sim-to-Real" gap continues to be a major obstacle due to visual and dynamic domain shifts, though techniques like domain randomization and prompt learning offer potential solutions. Additionally, there is a critical need to improve computational efficiency and real-time control performance through model compression and knowledge distillation to enable deployment on resource-constrained platforms. Future research should focus on the deep integration of cross-modal information, particularly combining tactile feedback with vision to enhance precision when handling objects with diverse textures or under occlusion. Furthermore, the scope of grasping tasks must be expanded to include more challenging scenarios such as manipulating flexible or transparent objects, coordinating multi-object interactions, and performing mobile manipulation. The ultimate goal is to fuse embodied foundation models with high-degree-of-freedom dexterous manipulation, injecting common-sense reasoning and task understanding into low-level control strategies to push robotic systems toward real-world application in daily life.