Grasp Stability Prediction for a Dexterous Robotic Hand Combining Depth Vision and Haptic Bayesian Exploration

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous robotic manipulation in unstructured environments requires reliable grasp stability prediction for unknown objects.
• Real-time local exploration is often needed to overcome inaccuracies in object modeling, even with prior information.

Purpose of the Study:

• To develop and evaluate an approach for predicting safe grasps of unknown objects using depth vision and tactile feedback.
• To compare probabilistic methods, specifically Bayesian Optimization, against random exploration for grasp planning.

Main Methods:

• Object pose estimation using RGB-D sensing.
• Haptic exploration to optimize grasp metrics.
• Comparison of standard Bayesian Optimization, unscented Bayesian Optimization, and uniform grid search.

Main Results:

• Probabilistic methods provide confident grasp predictions after limited exploratory observations.
• Unscented Bayesian Optimization identifies safer grasps by considering uncertainties in sensing and execution.
• Bayesian Optimization significantly outperforms random exploration.

Conclusions:

• Probabilistic methods, particularly unscented Bayesian Optimization, are effective for safe grasp prediction of unknown objects.
• The approach enhances robotic manipulation capabilities in complex, unpredictable environments.
• Integrating tactile feedback and vision improves grasp reliability and safety.