Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SANGO: Socially Aware Navigation through Grouped Obstacles

Rahath Malladi, Amol Harsh, Arshia Sangwan, Sunita Chauhan, Sandeep Manjanna

TL;DR

SANGO introduces socially aware navigation by dynamically grouping obstacles with DBSCAN and learning a policy via PPO. The method leverages a custom two-environment simulation suite (MOSANG and COG), a grouped-reward structure, and a 2D LiDAR-like perception to promote safe, socially compliant paths. Empirical results show significant improvements in discomfort reduction, collision avoidance, and successful navigation in crowded settings, highlighting practical potential for real-world service robots. Limitations include 2D motion modeling and the need to incorporate 3D trajectories and richer social norms for broader generalization.

Abstract

This paper introduces SANGO (Socially Aware Navigation through Grouped Obstacles), a novel method that ensures socially appropriate behavior by dynamically grouping obstacles and adhering to social norms. Using deep reinforcement learning, SANGO trains agents to navigate complex environments leveraging the DBSCAN algorithm for obstacle clustering and Proximal Policy Optimization (PPO) for path planning. The proposed approach improves safety and social compliance by maintaining appropriate distances and reducing collision rates. Extensive experiments conducted in custom simulation environments demonstrate SANGO's superior performance in significantly reducing discomfort (by up to 83.5%), reducing collision rates (by up to 29.4%) and achieving higher successful navigation in dynamic and crowded scenarios. These findings highlight the potential of SANGO for real-world applications, paving the way for advanced socially adept robotic navigation systems.

SANGO: Socially Aware Navigation through Grouped Obstacles

TL;DR

SANGO introduces socially aware navigation by dynamically grouping obstacles with DBSCAN and learning a policy via PPO. The method leverages a custom two-environment simulation suite (MOSANG and COG), a grouped-reward structure, and a 2D LiDAR-like perception to promote safe, socially compliant paths. Empirical results show significant improvements in discomfort reduction, collision avoidance, and successful navigation in crowded settings, highlighting practical potential for real-world service robots. Limitations include 2D motion modeling and the need to incorporate 3D trajectories and richer social norms for broader generalization.

Abstract

This paper introduces SANGO (Socially Aware Navigation through Grouped Obstacles), a novel method that ensures socially appropriate behavior by dynamically grouping obstacles and adhering to social norms. Using deep reinforcement learning, SANGO trains agents to navigate complex environments leveraging the DBSCAN algorithm for obstacle clustering and Proximal Policy Optimization (PPO) for path planning. The proposed approach improves safety and social compliance by maintaining appropriate distances and reducing collision rates. Extensive experiments conducted in custom simulation environments demonstrate SANGO's superior performance in significantly reducing discomfort (by up to 83.5%), reducing collision rates (by up to 29.4%) and achieving higher successful navigation in dynamic and crowded scenarios. These findings highlight the potential of SANGO for real-world applications, paving the way for advanced socially adept robotic navigation systems.

Paper Structure

This paper contains 15 sections, 1 equation, 7 figures, 1 table.

Table of Contents

  1. INTRODUCTION
  2. METHODOLOGY
  3. Grouping Reward, DBSCAN and PPO Algorithms
  4. Agent Architecture
  5. Reward Function
  6. ENVIRONMENTS & EXPERIMENTS
  7. MOSANG - Modular Socially Aware Navigation Gym
  8. COG - Complex Open Gym
  9. Learning Scenarios
  10. Environment and Obstacle Initialization
  11. Trajectory Generation
  12. Metrics
  13. Compute Setup Used
  14. RESULTS
  15. CONCLUSION

Figures (7)

  • Figure 1: An illustration of the proposed social navigation technique (SANGO). The agent groups the dynamic obstacles (humans in this case) into groups and uses this grouping to plan a path to the goal location.
  • Figure 2: Example scenarios with dynamic obstacles that can be used to train agents in diverse scenarios. Magenta circles represent the obstacles, the green star represents the robot, and the red cross represents the goal.
  • Figure 3: An Experiment Scenario: SANGO Agent (green star) navigating a COG environment with 40 static and 50 dynamic obstacles towards the goal (red cross). Cyan groups represent obstacles currently in the agent's vicinity, while the dark green group was nearby in the previous time step.
  • Figure 4: Discomfort Reduction Progression: Discomfort Scores achieved by SANGO agents over 580K training steps in different environments.
  • Figure 5: Group intrusion Progression: Group intrusion Rate achieved by SANGO agents over 580K training steps in diverse environments
  • ...and 2 more figures