Symbolic-Based Recognition of Contact States for Learning Assembly Skills

Area of Science:

• Robotics
• Machine Learning
• Human-Robot Interaction

Background:

• Imitation learning enables robots to acquire skills from human demonstrations, particularly useful for industrial assembly tasks.
• Recognizing contact states (CS) is crucial for successful assembly, typically relying on force/torque (F/T) data, which can be noisy and ambiguous.
• Existing methods face challenges due to task variations and signal interpretation difficulties in F/T data.

Purpose of the Study:

• To develop a robust method for recognizing contact states (CS) during assembly with geometrical variations.
• To improve the accuracy and reduce the computational load of CS recognition in robotic imitation learning.
• To investigate the use of symbolic and probabilistic approaches for interpreting force information.

Main Methods:

• Collected force/torque (F/T) data from human assembly trials.
• Pre-processed, segmented, and symbolized the F/T data.
• Trained a probabilistic model using the symbolic representations.
• Validated the model with unseen datasets.

Main Results:

• The probabilistic model successfully recognized contact states (CS) using only force information.
• The symbolic and probabilistic approach demonstrated improved recognition accuracy.
• The method showed reduced computational effort compared to traditional F/T analysis.

Conclusions:

• The developed model can effectively recognize contact states (CS) during assembly tasks, even with geometrical variations.
• This approach shows significant potential to aid imitation learning in robotics by providing reliable CS recognition.
• The findings suggest that force information, when processed symbolically and probabilistically, is sufficient for accurate CS identification.