Representing, learning, and controlling complex object interactions

Area of Science:

• Robotics
• Artificial Intelligence
• Control Theory

Background:

• Robots often face scenarios requiring indirect control of target objects via intermediate mechanisms.
• Reasoning about multi-object relationships is crucial for complex manipulation tasks.

Purpose of the Study:

• To develop a framework for representing and learning complex object interactions for robots.
• To enable robots to achieve indirect control over objects through intermediate interfaces.

Main Methods:

• Formalizing interactions as chains and graphs of Markov decision processes (MDPs).
• Learning interaction models from data.
• Collapsing complex models into a single MDP for optimal policy computation.
• Developing efficient planning algorithms for large state spaces.

Main Results:

• Demonstrated the ability to learn models of indirect object manipulation.
• Showcased a method to solve complex interaction systems by reducing them to a single MDP.
• Introduced a planning algorithm for efficient policy generation in large-scale systems.

Conclusions:

• The proposed framework effectively models and controls complex indirect robotic interactions.
• Learning from demonstration enables robots to master tasks requiring manipulation through intermediate objects.
• This approach is applicable to diverse tasks, including game control and operating dispensers.