New DDPM bimanual robot grasps large objects from single-view

Bimanual dexterous grasping of large objects remains a major challenge in robotics—most prior work relies on sequential manipulation or simulation. A new arXiv paper (2608.10383) from researchers including Ziming Li and Jiaqi Zhang proposes a real-world framework that enables a dual-arm robot to cooperatively grasp large objects using only single-view observations. The method uses a Denoising Diffusion Probabilistic Model (DDPM) to generate joint-level grasp configurations directly from segmented point clouds, bypassing the need for full 3D object models. This makes the system practical for unstructured environments where complete object geometry is unavailable.

The framework is built on a multimodal dataset capturing joint angles, visual observations, and force signals during bimanual manipulation. During execution, motion planning is combined with online grasp refinement to ensure physical stability and feasibility in real time. Experiments on a dual-arm robot show high success rates across unseen objects with diverse geometries and poses, and ablation studies confirm the importance of each system component. This work marks a step toward robust, human-like bimanual manipulation in real-world settings, with potential applications in warehouse automation, assistive robotics, and industrial material handling.