Physics simulation based downstream filtering of neural network grasp pose hypotheses of unknown objects

Kurzfassung

This work investigates physics simulation-based grasp evaluation as an alternative to ContactGraspNet’s (CGN) native grasp ranking for unknown objects. A pipeline was developed for grasping unknown objects, combining multi-camera and depth sensing, point cloud processing, shape completion, grasp generation, and physics-based evaluation using four grasp quality metrics. The study compared three evaluation methods: DrakeStatic, which combines MIT Drake physics simulation with grasp quality metrics; DrakeDynamic, which measures grasp stability through dynamic force application; and AltCGNOnlyStatic, which calculates contact forces using the mesh geometry and the generated grasp from CGN only. Results showed that physics-based evaluation (DrakeStatic) improved upon CGN’s native rankings in both best grasp selection and overall grasp quality correlation. However, DrakeStatic’s advantage over AltCGNOnlyStatic was smaller than expected. While DrakeStatic showed higher correlation with ground truth measurements, AltCGNOnlyStatic performed better at identifying single best grasps per object. DrakeDynamic, initially developed as an evaluation method, became the ground truth standard for the experiments. The experiment’s main limitation was the edge bleeding removal process, which created unwanted differences in simulation setup and asymmetric grasp filtering across objects. Additional limitations included shape completion artefacts and limited dataset size. The finding that DrakeStatic’s advantage over AltCGNOnlyStatic was smaller than expected and Drake simulation’s computational requirements, question the practical value of physics simulation for grasp evaluation.