Collective Behavior Clone with Visual Attention via Neural Interaction Graph Prediction

Kai Li, Zhaolong Ma, Liang Li, Shiyu Zhao

TL;DR

This paper introduces Collective Behavioral Cloning (CBC), a framework that learns both the hidden interaction mechanisms and control policies of robot swarms from trajectory data. By using a Graph Variational Autoencoder (GVAE) to predict interaction graphs and then applying behavioral cloning, CBC achieves more accurate prediction of swarm behavior and interaction patterns, demonstrated on a real-world decentralized robot swarm.

What Problem Does This Paper Address?

Understanding and replicating collective behaviors in swarm systems, like flocks of birds or schools of fish, is a significant challenge. Traditional approaches often rely on human-designed models for neighbor selection and motion control, which may not accurately reflect the true underlying interaction dynamics. The core problem is learning these hidden interaction mechanisms and control policies directly from observed swarm trajectories, especially in decentralized systems where robots lack direct communication. This paper aims to bridge this gap by developing a method that can infer these complex relationships and replicate expert swarm behavior.

Methods

The proposed Collective Behavioral Cloning (CBC) framework has two main phases:

  1. Learning the Local Interaction Graph: A Graph Variational Autoencoder (GVAE) is employed. The GVAE takes swarm trajectory data (positions and velocities) as input. It uses a multi-layer perceptron (MLP) and graph convolutional layers to embed node information, followed by an attention mechanism to model potential interactions between agents. An LSTM then extracts time-specific features. The output is a probability distribution over potential interaction graph edges for each agent at each time step. This phase essentially learns "who is talking to whom" in the swarm.
  2. Learning the Control Policy: Once the interaction graph is learned, behavioral cloning is used. A Multi-Layer Perceptron (MLP) acts as the action network, taking the states (relative positions and velocities) of the identified interacting neighbors as input and outputting the control action (velocity and angular velocity) for the agent. The loss function is the mean squared error between the predicted action and the expert's actual action.

For the real-world deployment, a visual attention network is trained. This network, combined with an object detection network (YOLOv5), acts as a proxy for the learned interaction graph. It allows decentralized robots to select interaction targets online based on visual cues, mimicking the learned interaction patterns without explicit communication.

Datasets

The paper uses a custom benchmark dataset generated from simulations of robot swarms. The simulation involves a leader robot following a predefined path (circular, linear, U-turn) while other agents follow, forming flocking behavior. The dataset consists of swarm trajectories with varying numbers of agents (3-9). For the real-world experiment, data was collected from a vision-based robot swarm system.

Key Insights and Ideas

  • Learning interactions is crucial: The paper strongly suggests that accurately replicating swarm behavior hinges on understanding and learning the underlying interaction mechanisms, rather than relying solely on predefined rules.
  • GVAE for interaction graphs: The GVAE proves effective at inferring complex, dynamic interaction graphs from trajectory data, outperforming other relational inference methods.
  • Visual attention as a communication proxy: In decentralized, communication-denied environments, a visual attention network can effectively learn to select interaction partners, mimicking the learned interaction graph and enabling robust collective behavior.
  • CBC bridges the gap: The CBC framework successfully combines interaction graph learning with behavioral cloning to replicate expert swarm behaviors, showing improved accuracy in both prediction and execution.

Key Results

  • Interaction Graph Prediction: The proposed GVAE model achieved a significantly higher F1 score (91.20%) for interaction graph edge prediction compared to state-of-the-art methods like dNRI (61.25%) on simulated swarm data. This performance was consistent across different numbers of agents.
  • Collective Behavioral Cloning: In real-world experiments, the CBC framework demonstrated superior performance in replicating expert swarm behavior, achieving the smallest action error (AE), trajectory error (TE), and average mutual distance (AMD) compared to vision-only, range-based, and K-nearest neighbor selection methods. For instance, AE was 0.028 ± 0.016, TE was 0.95 ± 0.40, and AMD was 0.85 ± 0.31.
  • Robustness: The framework showed robustness in maintaining desired relative distances and coherent behavior, even in scenarios with multiple groups of robots.

Limitations and Future Work

  • Scalability: While the GVAE performed well across different agent numbers in simulation, the paper doesn't extensively discuss the computational scalability of the full CBC framework for very large swarms in real-world scenarios.
  • Generalizability: The visual attention network is trained based on the GVAE's predictions. Its effectiveness might depend on the accuracy of the GVAE's learned graph and the quality of the generated training data.
  • Future Work: The authors suggest extending this framework to study complex behaviors in humans or natural swarms, implying potential applications beyond robotics. Further research could explore more sophisticated interaction modeling techniques or investigate how to adapt the learned policies to dynamic environmental changes.

Summary updated

Abstract

In this paper, we propose a framework, collective behavioral cloning (CBC), to learn the underlying interaction mechanism and control policy of a swarm system. Given the trajectory data of a swarm system, we propose a graph variational autoencoder (GVAE) to learn the local interaction graph. Based on the interaction graph and swarm trajectory, we use behavioral cloning to learn the control policy of the swarm system. To demonstrate the practicality of CBC, we deploy it on a real-world decentralized vision-based robot swarm system. A visual attention network is trained based on the learned interaction graph for online neighbor selection. Experimental results show that our method outperforms previous approaches in predicting both the interaction graph and swarm actions with higher accuracy. This work offers a promising approach for understanding interaction mechanisms and swarm dynamics in future swarm robotics research. Code and data are available <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">6</sup>.
BibTeX
@misc{li2025collective,
  author = {Kai Li and Zhaolong Ma and Liang Li and Shiyu Zhao},
  title = {{Collective Behavior Clone with Visual Attention via Neural Interaction Graph Prediction}},
  year = {2025},
  eprint = {2503.06869},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2503.06869},
}
Figure