Beyond imitation: Zero-shot task transfer on robots by learning concepts as cognitive programs.

Area of Science:

• Robotics
• Artificial Intelligence
• Cognitive Science

Background:

• Humans excel at inferring concepts from visual data and applying them across diverse physical contexts.
• Current robots struggle with high-level concept representation, limiting intent understanding and task transfer capabilities.

Purpose of the Study:

• To develop a computational framework for robots to learn and infer high-level concepts, enhancing their ability to understand human intent and generalize tasks.
• To bridge the gap between human-like concept learning and robotic capabilities.

Main Methods:

• A novel computational framework representing concepts as programs on a cognitive computer architecture.
• The architecture includes systems for visual perception, working memory, and action control with a specialized instruction set.
• Concept inference is achieved by inducing programs that map visual input to desired output, incorporating imagination and recursion.

Main Results:

• Demonstrated a robot's ability to interpret novel concepts from schematic images using the framework.
• Showcased the robot's successful application of inferred concepts in significantly different environments.
• Established a hierarchical learning process where prior concepts facilitate learning of more complex ones.

Conclusions:

• The developed framework enables robots to learn abstract concepts, mirroring human cognitive abilities.
• Integrating cognitive science principles into machine learning advances robots with interpretable representations and common sense reasoning.
• This approach moves closer to creating robots that can understand and execute tasks with greater flexibility and intelligence.