Learning Inverse Statics Models Efficiently With Symmetry-Based Exploration.

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Learning robot kinematics and dynamics models is computationally expensive and time-consuming.
• Sampling the entire robot configuration space is often intractable.
• Efficient learning of inverse statics models, crucial for gravity compensation, remains a challenge.

Purpose of the Study:

• To develop an efficient approach for learning inverse statics models by exploring limited configuration space.
• To leverage symmetry properties of inverse statics mappings to reduce sample complexity.
• To demonstrate the general applicability of symmetry exploitation for both online and offline learning algorithms.

Main Methods:

• Exploiting symmetric configurations that require identical absolute motor torques.
• Discovering symmetric configurations and learning functional relations between them.
• Generating multiple training samples from single sampled configuration-torque pairs.
• Applying modified Direction Sampling for online learning and lattice sampling for offline learning.

Main Results:

• Reduced the number of samples required for learning inverse statics models significantly.
• Successfully learned inverse statics mappings for the entire configuration space of a 2R planar robot and a 3R simplified human arm.
• Achieved sample reduction factors of approximately 8 for 2R and 16 for 3R manipulators.

Conclusions:

• Exploiting symmetries is an effective strategy for efficient learning of inverse statics models.
• The proposed method drastically reduces sample requirements, making robot model learning more practical.
• The approach is applicable to various learning algorithms and robot manipulators.