This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
This paper proposes a novel volumetric grip model that is equivariant with respect to vertical rotation, significantly improving sampling efficiency. It utilizes a trihedral volumetric feature representation that projects 3D features onto three canonical planes. We introduce a novel trihedral feature design in which features in the horizontal plane are equivariant with respect to 90-degree rotations, while the sum of features in the other two planes is invariant with respect to the same transformation. This design is achieved through a novel deformable steerable convolution that combines the adaptability of deformable convolutions with the rotational isovariance of steerable convolutions. This allows the receptive field to adapt to local object geometry while maintaining the isovariance property. Furthermore, we develop isovariance adaptation in GIGA, a state-of-the-art volumetric grip planner, and IGD. Specifically, we derive a new isovariance formulation of IGD's deformable attention mechanism and propose an isovariance generation model for grip orientation based on flow matching. We provide a detailed analytical justification for the proposed isovariance property and validate the approach through extensive simulations and field experiments. The results demonstrate that the proposed projection-based design significantly reduces computational and memory costs. Furthermore, the isosceles grip model built on triangular features consistently outperforms the anisoceles model, achieving higher performance with less computational overhead. The video and code can be found at https://mousecpn.github.io/evg-page/ .
A novel volume grip model is presented that significantly improves sample efficiency by exploiting isotropy for vertical axis rotation.
◦
A proposal for a three-sided volume feature representation that efficiently reduces computational and memory costs.
◦
Development of an isosceles adaptation method for state-of-the-art grip planners such as GIGA and IGD.
◦
Performance verification and excellence confirmation through simulation and actual experiments.
•
Limitations:
◦
The performance of the proposed model may be limited to specific environments (simulated and specific real-world environments). Further research is needed to determine its generalization performance across diverse environments.
◦
Further research is needed to determine whether the three-dimensional feature representation is optimal for all types of objects.
◦
Further analysis is needed to determine the resistance to noise and uncertainty that may arise in real-world applications.
