Efficient multitask learning with an embodied predictive model for door opening and entry with whole-body control.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Robots require robust, adaptable models for human-like tasks, but development and maintenance are costly.
• Deep reinforcement learning offers automated model acquisition but incurs high real-world learning costs due to extensive trial and error.

Purpose of the Study:

• To develop a method for robots to perform complex real-world tasks with reduced design and teaching costs.
• To enable robots to generate appropriate motions and adapt to environmental changes autonomously.

Main Methods:

• A prediction error minimization principle was employed.
• A module integration method was devised, incorporating a mechanism to switch between modules based on prediction error.
• Real-time prediction error calculation allowed for dynamic module selection and task sequencing.

Main Results:

• The robot successfully generated appropriate motions for tasks like door opening, adapting to variations in position, color, and pattern.
• A sequence of tasks, including opening a door and passing through, was executed by linking multiple modules.
• The method demonstrated effective autonomous operation and responsiveness to sudden environmental and procedural changes.

Conclusions:

• The proposed method significantly lowers design and teaching costs for complex robotic tasks.
• Robots can autonomously adapt and execute sequential tasks in dynamic real-world environments.
• This approach enhances robot adaptability and efficiency in practical applications.