Learning efficient haptic shape exploration with a rigid tactile sensor array.

Area of Science:

• Robotics
• Artificial Intelligence
• Sensory Perception

Background:

• Haptic exploration is crucial for robots and humans to interact with and recognize objects.
• Robotic haptic exploration lags behind robot vision due to limited models, sensors, and training data.
• Human haptic exploration involves sophisticated sensory-motor skills and exploratory procedures.

Purpose of the Study:

• To develop a novel architecture for learning generative models of haptic exploration in robots.
• To integrate advances in recurrent visual attention models with human haptic search strategies.
• To create a system capable of learning haptic object identification through simulated exploration.

Main Methods:

• Developed a Haptic Attention Model (HAM) architecture in a 3D simulated environment.
• Utilized a reinforcement learning framework inspired by the Recurrent Attention Model.
• Employed a multi-module neural network for feature extraction and sequential data integration, optimizing perception-action loops.

Main Results:

• The HAM successfully learned haptic object identification for rigid objects with 1D shape features.
• Achieved near 100% accuracy in object contour exploration optimized for a 16x16 tactile sensor array.
• Demonstrated effective simultaneous optimization of feature extraction, temporal integration, and control strategy.

Conclusions:

• The proposed Haptic Attention Model significantly advances robotic haptic exploration capabilities.
• The approach bridges the gap between human sensory-motor skills and robotic tactile perception.
• This work paves the way for more sophisticated object discrimination and manipulation in robots.