Reproduce “Learning Social Heuristics for Human-Aware Path Planning” on OpenResearch

About this paper

Andrea Eirale, Matteo Leonetti, Marcello Chiaberge

Researchers from Politecnico di Torino and King's College London developed Heuristic Planning with Learned Social Value (HPLSV), a method that integrates complex, activity-dependent social norms into classical path planning algorithms. By learning a social heuristic offline through Reinforcement Learning, the robot demonstrates socially acceptable navigation, such as joining a queue without cutting, in both discrete and continuous environments.

Key insights

• An ego-centric state representation ensures the learned social heuristic is generalizable across different environments and scales, without requiring retraining for every new scene.
• Decoupling the social behavior learning from the core navigation component allows for efficient integration into robust classical planners, adding social intelligence with minimal computational overhead.
• The social heuristic can be selectively activated using a confidence threshold, ensuring it only influences planning when contextually relevant, thus maintaining efficiency in general navigation.

Summary by alphaXiv.