Machine learning-driven self-discovery of the robot body morphology

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Traditional robot morphology is assumed known, with external sensors used for calibration.
• Learning robot morphology autonomously remains an underexplored area.

Purpose of the Study:

• To investigate if robots can learn their own morphology using only proprioceptive signals.
• To develop a method for inferring robot morphology from internal sensor data.

Main Methods:

• Proposed a mutual information-based representation called proprioceptive information graphs (π-graphs).
• Analyzed pairwise signal relationships within π-graphs to identify kinematic principles.
• Inferred mechanical topology and kinematic descriptions (joint locations/orientations).

Main Results:

• Successfully inferred robot morphology from π-graphs across diverse robotic systems (manipulator, hexapod, humanoid).
• Demonstrated effective inference both offline and online, independent of robot complexity.
• Validated the agent-centric approach for morphological self-discovery.

Conclusions:

• Proprioceptive information graphs enable robots to autonomously learn their morphology.
• This method reduces reliance on external measurement devices for robot calibration and understanding.
• The findings open new avenues for adaptable and self-aware robotic systems.