Simulated mental imagery for robotic task planning.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Traditional AI planning in robotics relies on labor-intensive symbolic domain descriptions.
• Human task planning often utilizes intuitive mental imagery, a capability lacking in current AI.
• Sub-symbolic AI methods like deep reinforcement learning offer alternatives but lack interpretability.

Purpose of the Study:

• To introduce a novel sub-symbolic planning method for robots inspired by human mental imagery.
• To enable robots to generate human-interpretable action plans without explicit symbolic domain knowledge.
• To explore the efficacy of combining convolutional neural networks and generative adversarial networks for robotic planning.

Main Methods:

• Proposed Simulated Mental Imagery for Planning (SiMIP), a sub-symbolic approach.
• SiMIP integrates perception, simulated action, success checking, and re-planning on 'imagined' images.
• Utilized a combination of convolutional neural networks and generative adversarial networks for implementation.

Main Results:

• Demonstrated algorithmic soundness of mental imagery-based planning for robots.
• Enabled robots to generate action plans directly from existing scenes without symbolic descriptions.
• Plans generated by SiMIP were found to be human-interpretable, unlike deep reinforcement learning approaches.

Conclusions:

• Simulated Mental Imagery for Planning (SiMIP) offers a viable sub-symbolic alternative for robotic task planning.
• The method allows for intuitive, human-interpretable planning, bridging the gap between human and robot task execution.
• Future work can quantify the efficiency and success rate of SiMIP on various robotic tasks, such as object packing.