A Hierarchical VLM-to-TD3 Framework with Novel Object Coordinate Estimation and Persistent Spatial Memory for...

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Autonomous indoor navigation faces challenges in cluttered environments, particularly for long-horizon tasks requiring object recognition and spatial reasoning.
• Deep reinforcement learning (DRL) methods like TD3 excel at obstacle avoidance but struggle with semantic understanding and memory-based navigation.
• Existing approaches often lack seamless integration of vision-language understanding with navigation control.

Purpose of the Study:

• To present a novel, lightweight, hierarchical two-stage framework for language-conditioned autonomous indoor navigation.
• To integrate a locally deployed vision-language model (VLM), semantic object coordinate memory, and a TD3-based DRL controller.
• To improve navigation efficiency and goal-reaching performance in complex indoor settings.

Main Methods:

• Stage 1: Semantic exploration using odometry, 2D LiDAR, and VLM for object recognition, building a geometric map and storing object coordinates in a JSON semantic memory.
• Stage 2: Language query retrieves target coordinates from memory, feeding into a TD3 policy for mapless navigation using odometry and RGB-D perception.
• The framework unifies VLM-based coordinate estimation, LiDAR mapping, RGB-D perception, and language grounding.

Main Results:

• The proposed framework demonstrated consistent goal-reaching performance and improved navigation efficiency in ROS-integrated simulations compared to baseline methods.
• The VLM-based recognition module showed more reliable semantic grounding and coordinate estimation than YOLO-World and Grounding DINO, especially for flexible queries.
• The system successfully navigated using natural language commands and object memory.

Conclusions:

• The developed hierarchical framework effectively integrates VLM, semantic memory, and DRL for robust language-conditioned indoor navigation.
• This approach offers a significant advancement in handling complex indoor environments and natural language instructions for robots.
• The VLM-based semantic memory and navigation policy provide a reliable solution for open-vocabulary indoor navigation tasks.