Multi Agent Navigation in Unconstrained Environments using a Centralized Attention based Graphical Neural Network Controller

TL;DR

Experimental results demonstrate that the model is powerful enough to generalize even to situations with more vehicles than in the training data, and outperforms comparable graphical neural network architectures.

Abstract

In this work, we propose a learning based neural model that provides both the longitudinal and lateral control commands to simultaneously navigate multiple vehicles. The goal is to ensure that each vehicle reaches a desired target state without colliding with any other vehicle or obstacle in an unconstrained environment. The model utilizes an attention based Graphical Neural Network paradigm that takes into consideration the state of all the surrounding vehicles to make an informed decision. This allows each vehicle to smoothly reach its destination while also evading collision with the other agents. The data and corresponding labels for training such a network is obtained using an optimization based procedure. Experimental results demonstrates that our model is powerful enough to generalize even to situations with more vehicles than in the training data. Our method also outperforms comparable graphical neural network architectures. Project page which includes the code and supplementary information can be found at https://yininghase.github.io/multi-agent-control/

Figures (5)

• Figure 1: Overview of multi-agent control: The initial configuration (on the left) shows five vehicles colored black, red, green, blue and olive. The initial starting state of the vehicles is represented by a rectangle with solid boundaries. The arrow within each rectangle depicts the corresponding orientation of that vehicle. Meanwhile, the rectangles with broken boundaries represents the desired destination/target position of each vehicle. We would like to produce the sequence of control actions such that the five vehicles safely reach their destination state without colliding with each other or the circled obstacle. These control actions are produced by the Attention Based Graphical Neural Network (A-GNN). The A-GNN receives information about the current state and desired destination state of all the five vehicles along with information about any obstacle. The network outputs the control commands for all the five vehicles together. These control commands are executed for all the five vehicles simultaneously. Each vehicle then attains a new state. This new state is then fed again to the A-GNN as the current state to predict the new steering command. This process is iteratively repeated until all the vehicles reach their corresponding destination state. The trajectory traversed by all the vehicles as a result of this iterative process is shown on the right of the figure. Some video examples of our model demonstrating this can be found: https://yininghase.github.io/multi-agent-control#Results-of-our-Model
• Figure 2: Shows the trajectory traversed by the vehicles as a result of applying the control commands predicted by the five different models for a scenario containing 6 vehicles and 3 obstacles. The red dots on the trajectories show the point of collision between the vehicles. Except for our model, all other models have plenty of collisions. Video can be found: https://yininghase.github.io/multi-agent-control#Comparison-with-Other-Models.
• Figure 3: An example of a conservative optimization yielding a sample trajectory wherein the vehicles halt rather than take the risk of collision in an attempt to reach their destination by trying to pass between the two obstacles.
• Figure 4: Predicted Trajectories of our model with Attention Logits: The color of pixel at row $i$ column $j$ shows the attention logits of object $i$ on object $j$. The lower the logits value is, the darker is the color and vice versa. The gray value on the diagonal means that a vehicle is neutral to attending to itself. Corresponding video can be found in the codebase: https://yininghase.github.io/multi-agent-control#Attention-Mechanism-of-our-Model
• Figure 5: Schematics of the U-Attention Block