Adaptive Prior Selection for Repertoire-Based Online Adaptation in Robotics

Area of Science:

• Robotics
• Machine Learning
• Artificial Intelligence

Background:

• Repertoire-based learning offers data-efficient robot adaptation by selecting pre-learned policies.
• Previous methods assumed a single repertoire is sufficient, limiting adaptation to diverse situations.

Purpose of the Study:

• To develop a more robust adaptation method by generating and selecting from multiple situation-specific repertoires.
• To introduce the Adaptive Prior selection for Repertoire-based Online Learning (APROL) algorithm for planning actions with unknown situational priors.

Main Methods:

• Generated multiple repertoires for various robot situations (e.g., damage, different environments).
• Developed APROL to select the most relevant prior policy based on the current situation.
• Evaluated APROL on simulated robotic arm object pushing and hexapod robot goal reaching tasks.
• Compared APROL against Reset-free Trial and Error (RTE) and single-repertoire baselines.

Main Results:

• APROL demonstrated superior performance by solving tasks in less interaction time compared to baselines.
• Successfully adapted a real, damaged hexapod robot to reach a goal while avoiding obstacles.
• Showcased the algorithm's ability to learn compensatory policies for effective online adaptation.

Conclusions:

• APROL significantly enhances repertoire-based learning by utilizing multiple situation-specific repertoires.
• The algorithm provides a more effective and data-efficient approach for robot adaptation in unknown or changing environments.
• APROL's successful real-world demonstration highlights its practical applicability for robust robotic systems.