GoalGrasp: Grasping Goals in Partially Occluded Scenarios without Grasp Training

TL;DR

GoalGrasp presents an object-level 6-DoF grasp pose detection method that enables user-specified target grasping without grasp-specific training and mitigates partial occlusion by leveraging 3D bounding boxes and simple human grasp priors. The approach uses Recursive Cross-View (RCV) for 3D detection without 3D annotations, then generates dense grasp poses per object with shape-specific heuristics, filtered for feasibility and scored with a novel stability metric. Experiments show dense grasp pose generation for 1,000 scenes across 18 objects, higher stability than competitive learning-based methods, and real-robot success rates up to 94% for user-specified grasps and 92% under partial occlusion. The work demonstrates practical target-oriented grasping without training and highlights avenues for expanding object coverage, automating heuristics, and integrating obstacle avoidance for robust service-robot applications.

Abstract

Grasping user-specified objects is crucial for robotic assistants; however, most current 6-DoF grasp detection methods are object-agnostic, making it challenging to grasp specific targets from a scene. To achieve that, we present GoalGrasp, a simple yet effective 6-DoF robot grasp pose detection method that does not rely on grasp pose annotations and grasp training. By combining 3D bounding boxes and simple human grasp priors, our method introduces a novel paradigm for robot grasp pose detection. GoalGrasp's novelty is its swift grasping of user-specified objects and partial mitigation of occlusion issues. The experimental evaluation involves 18 common objects categorized into 7 classes. Our method generates dense grasp poses for 1000 scenes. We compare our method's grasp poses to existing approaches using a novel stability metric, demonstrating significantly higher grasp pose stability. In user-specified robot grasping tests, our method achieves a 94% success rate, and 92% under partial occlusion.

Figures (9)

• Figure 1: 2D labeling and inferring 3D bounding boxes for RCV on the collected data.
• Figure 2: 3D-TABLETOP-OBJECT dataset with 15 categories: large box, small box, large cylinder, small cylinder, bowl, cucumber, banana, tape, screw, apple, lemon, grapefruit, pen, jar, mug.
• Figure 3: Grasp poses generation for box-shaped objects. (a) Sampling trajectory (ST) for grasp points. (b) Generated grasp poses without filtering.
• Figure 4: The illustration of the novel stability metric.
• Figure 5: The illustration of overcoming occlusion.