Socially Adaptive Path Planning Based on Generative Adversarial Network

Yao Wang, Jiajie Yu, Yuqi Kong, Lining Sun, Chao Liu, Jiankun Wang, Wenzheng ChiIEEE Transactions on Intelligent Vehicles 2024 · DOI

TL;DR

This paper introduces GAN-RTIRL, a novel framework for socially adaptive robot path planning that leverages Generative Adversarial Networks (GANs) to learn human-like navigation behaviors. By combining GANs with the RRT* algorithm and Inverse Reinforcement Learning (IRL), the system generates paths that are not only efficient but also prioritize pedestrian comfort and social norms, outperforming existing methods in complex human-robot interaction scenarios.

What Problem Does This Paper Address?

Autonomous robots navigating in human-populated environments face a critical challenge: how to interact with pedestrians in a way that is both safe and socially acceptable. Traditional path planning algorithms, like Dijkstra's, A*, and even RRT, treat pedestrians merely as moving obstacles. While they can find optimal paths in terms of distance or time, they often disregard the psychological comfort of humans, leading to interactions that feel unnatural or even unsettling. This can manifest as robots cutting too close to people, not anticipating their movements, or generally behaving in ways that humans wouldn't. The core problem is that these algorithms lack an understanding of "social rules" and human behavioral tendencies, which are crucial for seamless human-robot interaction. Existing learning-based methods have shown promise but often struggle with generalization across diverse and dynamic environments.

Methods

The paper proposes a multi-stage approach:

  1. GAN Model for Generalization: A Generative Adversarial Network (GAN) is trained to achieve strong generalization capabilities. This GAN is designed to learn a mapping from environmental features to a "cost value" for potential path nodes. This cost value aims to capture socially aware navigation preferences.
  2. GAN-RRT* Algorithm: This is the core path planning algorithm. It integrates the trained GAN model into the RRT* framework. Instead of relying solely on predefined cost functions (like distance), GAN-RRT* uses the GAN to evaluate the "social cost" of growing nodes. This allows the RRT* tree to explore paths that are not just geometrically optimal but also socially preferable. The GAN generates a cost for each node, and a discriminator helps refine this cost by distinguishing between "real" (demonstration-like) and "fake" (less socially aware) path segments.
  3. GAN-RTIRL Framework: This framework combines GAN-RRT* with Inverse Reinforcement Learning (IRL). The goal here is to further improve the "homotopy rate" – essentially, how similar the generated path is to a human-demonstrated path in terms of its overall shape and avoidance strategies. The IRL component uses demonstration paths (collected from human volunteers) to fine-tune the GAN model, making the robot's generated paths increasingly anthropomorphic. The framework iteratively trains the generator and discriminator, aiming for an equilibrium where the generated paths closely mimic human behavior.

Datasets

The paper mentions using "demonstration paths" generated by volunteers remotely controlling a robot. These paths are collected from various "scenarios" and "maps" to train and evaluate the GAN-RTIRL framework. The paper also describes a "scenario collection phase" where map, people, start, and goal information are integrated. For evaluation, they designed 100 static maps of size 324x257 pixels, splitting them into training (75 scenarios) and testing (25 scenarios) sets. Real-world experiments involved a robot named Xiaopang and 30 student participants for a questionnaire-based evaluation.

Key Insights and Ideas

  • GANs for Social Cost: Using GANs to learn a "social cost" for path nodes is a novel way to imbue robots with an understanding of human preferences beyond simple obstacle avoidance. This moves beyond hand-crafted cost functions.
  • Homotopy Rate as a Key Metric: The emphasis on improving the homotopy rate highlights the importance of generating paths that are not just safe but also intuitively "human-like" in their avoidance strategies.
  • Generalization is Crucial: The paper explicitly addresses the challenge of generalizing to diverse and changing human-robot interaction scenarios, a common limitation in prior work. The GAN's inherent generalization capabilities are key to this.
  • Turing-like Test for Anthropomorphism: The use of a "Turing-like test" in the evaluation, where participants try to distinguish between robot-generated and human-demonstrated paths, provides a compelling qualitative measure of the algorithm's success in achieving anthropomorphic behavior.

Key Results

  • Improved Homotopy Rate: The GAN-RRT* method achieved a homotopy rate of 94% in simulations, significantly higher than RRT* (75%) and NN-RRT* (90%).
  • Reduced Dissimilarity and Feature Difference: In simulations, GAN-RRT* showed the lowest dissimilarity and feature difference compared to RRT* and NN-RRT*, indicating that its generated paths are more similar to demonstration paths.
  • Higher Anthropomorphic Degree: Experimental results, including average Likert scale scores from participants, indicated that GAN-RRT* generates paths that make pedestrians more comfortable and are perceived as more anthropomorphic.
  • Real-World Validation: A Turing-like test in real-world experiments showed an anthropomorphic confidence of 80.77% for the GAN-RRT* algorithm.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, the use of GANs and RRT* suggests a potentially higher computational cost compared to simpler algorithms, which might be a consideration for real-time applications on resource-constrained robots.
  • Scenario Diversity: Although the paper emphasizes generalization, the complexity and diversity of real-world human-robot interactions are vast. Further work could explore even more challenging and dynamic scenarios.
  • Human Behavior Modeling: The current approach learns from demonstration paths. Future work could explore more explicit modeling of human intent, prediction, and more complex social interactions. The paper suggests that the GAN model can be adapted to different network architectures, hinting at flexibility for future improvements.

Summary updated

Abstract

The natural interaction between robots and pedestrians in the process of autonomous navigation is crucial for the intelligent development of mobile robots, which requires robots to fully consider social rules and guarantee the psychological comfort of pedestrians. Learning-based socially adaptive algorithms have shown promise in specific human-robot interaction environments, yet generalization of robot path planning requires improvement. This work combines the generative adversarial network (GAN) with the Optimal Rapidly-exploring Random Tree (RRT*) navigation algorithm. It proposes a GAN model with strong generalization performance and a GAN-RRT* algorithm for path generation in human-robot interaction environments. Additionally, a socially adaptive path planning framework named GAN-RTIRL is presented, which combines GAN with Rapidly-exploring random Trees Inverse Reinforcement Learning (RTIRL) to enhance the homotopy rate between planned and demonstration paths. The GAN-RRT* path planner can update the GAN model from the demonstration path, enabling the robot to generate more anthropomorphic paths and have stronger generalization in complex environments. Experimental results indicate that the proposed method effectively improves the anthropomorphic degree of robot motion planning and the homotopy rate.
BibTeX
@article{wang2024socially,
  author = {Yao Wang and Jiajie Yu and Yuqi Kong and Lining Sun and Chao Liu and Jiankun Wang and Wenzheng Chi},
  title = {{Socially Adaptive Path Planning Based on Generative Adversarial Network}},
  journal = {IEEE Transactions on Intelligent Vehicles},
  year = {2024},
  doi = {10.1109/tiv.2024.3478219},
}
Figure