Autonomous robots with socially-aware navigation using memory-assisted deep reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Human-Robot Interaction

Background:

• Service robot success in human environments depends on flexible navigation.
• Human stochasticity and dynamism in crowded spaces challenge robot navigation.
• Existing methods struggle with dynamic human behavior.

Purpose of the Study:

• To develop a robust navigation system for service robots in human-centric environments.
• To address challenges posed by dynamic human behavior using deep reinforcement learning.
• To enhance robot decision-making by prioritizing human interactions and safety.

Main Methods:

• Implemented a deep reinforcement learning approach named Adaptive Robot Safety Algorithm (ARSA).
• Incorporated bidirectional gated recurrent unit layers for long-term environmental memory.
• Integrated dynamic warning zones and prioritized human behaviors in the learned policy.

Main Results:

• ARSA policy improved success rate by 4% in simulations.
• Maintained a collision rate below 4% and reduced navigation time by 14%.
• Validated through real-world experiments, showing smooth, collision-free navigation.

Conclusions:

• The ARSA framework offers superior efficiency and safety for robot navigation in dynamic human environments.
• Memory-assisted deep reinforcement learning effectively handles environmental stochasticity.
• ARSA enables proactive and informed robot decision-making for safer human-robot interaction.