Machine Learning for Haptics: Inferring Multi-Contact Stimulation From Sparse Sensor Configuration.

Area of Science:

• Robotics
• Machine Learning
• Haptics

Background:

• Dexterous robots require robust haptic (touch) feedback systems.
• Current solutions are limited to small surface areas, leaving a gap for large, arbitrary 3D robot surfaces.
• Affordable and robust techniques for large-area 3D haptic sensing are needed.

Purpose of the Study:

• To introduce a general machine learning framework for inferring multi-contact haptic forces on a 3D robot limb surface.
• To predict surface deformation from sparse sensor data and identify unknown contact points and forces.
• To demonstrate the framework's effectiveness using transfer learning, even with limited single-contact training data.

Main Methods:

• A machine learning framework was developed to infer multi-contact haptic forces from internal deformation.
• The framework predicts surface deformation patterns from sparse physical sensor data.
• Transfer learning was employed, using single-contact data to train for multi-contact scenarios on a modified Poppy robot limb.

Main Results:

• The framework accurately infers the number, location, and force magnitude of multiple contact points.
• High accuracy was achieved even with only 10 strain-gauge sensors.
• The system demonstrated effectiveness for multi-contact scenarios using transfer learning from single-contact data.

Conclusions:

• The proposed machine learning framework provides an accurate method for multi-contact haptic force inference on 3D robot surfaces.
• The approach is versatile, applicable to arbitrary surfaces and sensor types, provided training data is available.
• This work advances robotic capabilities by enabling sophisticated touch sensing on large, complex robot bodies.