Mapless mobile robot navigation at the edge using self-supervised cognitive map learners

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Neuroscience

Background:

• Mobile agent navigation in unknown environments is crucial for autonomy.
• Deep Reinforcement Learning (DRL) shows promise but faces challenges in complexity, memory, and edge hardware adaptation.
• Current DRL methods often require intricate reward functions and extensive computational resources.

Purpose of the Study:

• To propose a brain-inspired, self-supervised navigation method for mobile agents.
• To overcome the limitations of complex DRL approaches for resource-constrained environments.
• To enable efficient navigation on edge neuromorphic processors.

Main Methods:

• A shallow neural network architecture trained with a local learning rule.
• Self-supervised learning combining novelty-based and random walks for exploration.
• Elimination of error backpropagation for reduced memory overhead.

Main Results:

• Achieved comparable goal-reaching accuracy and path length to state-of-the-art Deep Q Network (DQN).
• Utilized a similar or lower number of parameters, operations, and training iterations compared to DQN.
• Demonstrated reduced memory footprint and suitability for edge neuromorphic hardware.

Conclusions:

• The proposed brain-inspired method offers an efficient alternative for autonomous navigation.
• Self-supervised learning and shallow architectures enhance agent autonomy and adaptability.
• This approach facilitates the development of resource-efficient embodied neuromorphic agents.