DexGraspDiffuser: Target-Coupled Grasp and Action Diffusion for Dexterous Grasping

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Dexterous grasping with multi-finger robotic hands is crucial for general-purpose robotic manipulation but faces challenges in high-dimensional configurations and complex dynamics.
• Current methods often separate grasp target generation from execution policy learning, leading to inconsistencies and limiting overall performance.
• The need for integrated approaches that ensure consistency between grasp goals and downstream control is critical for advancing robotic manipulation.

Purpose of the Study:

• To propose DexGraspDiffuser, a novel framework that couples grasp target generation with action diffusion for enhanced dexterous grasping.
• To improve the consistency between generated grasp goals and the execution policy for more robust robotic manipulation.
• To achieve higher success rates and better generalization in robotic grasping tasks, especially with unseen objects.

Main Methods:

• Developed a two-stage diffusion framework: GraspDiffusion for generating diverse and physically plausible target grasps from object point clouds.
• Implemented a Goal-Conditioned Diffusion Policy to predict temporally coherent action sequences, conditioned on the target grasp and current observations.
• Utilized receding-horizon execution for action-prefix execution and online replanning to enhance inference robustness and adaptability.

Main Results:

• Achieved success rates of 0.76 on training objects, 0.72 on unseen objects from seen categories, and 0.68 on objects from unseen categories.
• Demonstrated a three-split average success rate of 0.72 with a low train-to-unseen generalization gap of 0.08.
• Outperformed the UniDexGrasp-T baseline by 3.3 percentage points in average success rate and reduced mean position error by 0.53 cm.

Conclusions:

• The target-coupled grasp and action diffusion approach significantly improves grasp quality and execution accuracy in dexterous robotic manipulation.
• DexGraspDiffuser demonstrates strong generalization capabilities, effectively handling objects not encountered during training.
• The framework contributes to enhanced closed-loop stability and robustness, paving the way for more capable general-purpose robotic systems.