|
1 | | -\page coverage-control-problem Problem Statement |
| 1 | +\page coverage-control-problem Theoretical Background |
2 | 2 | \tableofcontents |
3 | 3 |
|
| 4 | +# Coverage Control Problem |
4 | 5 | ## Introduction |
5 | 6 | Coverage control is the problem of navigating a robot swarm to collaboratively monitor features or a phenomenon of interest not known _a priori_. |
6 | 7 | The goal is to provide sensor coverage based on the importance of information at each point in the environment. |
@@ -63,3 +64,38 @@ In such a setting, a coverage control algorithm needs to provide the following b |
63 | 64 | Designing such decentralized algorithms is challenging and can be intractable for complex systems. |
64 | 65 | This motivates us to use a learning-based approach to design a decentralized coverage control algorithm. |
65 | 66 | The \ref lpac with GNN addresses the above challenges and provides a scalable and robust solution to the problem. |
| 67 | + |
| 68 | +------ |
| 69 | + |
| 70 | +# LPAC Architecture |
| 71 | + |
| 72 | +## Navigation of Robot Swarms |
| 73 | +Navigating a swarm of robots through an environment to achieve a common collaborative goal is a challenging problem, especially when the sensing and communication capabilities of the robots are limited. |
| 74 | +These problems require systems with high-fidelity algorithms comprising three key capabilities: perception, action, and communication, which are executed in a feedback loop, i.e., the Perception-Action-Communication (PAC) loop. |
| 75 | +To seamlessly scale the deployment of such systems across vast environments with large robot swarms, it is imperative to consider a decentralized system wherein each robot autonomously makes decisions, drawing upon its own observations and information received from neighboring robots. |
| 76 | + |
| 77 | +## The Challenge |
| 78 | +Designing a navigation algorithm for a decentralized system is challenging. |
| 79 | +The robots perform perception and action independently, while the communication module is the only component that can facilitate robot collaboration. |
| 80 | +Under limited communication capabilities, the robots must decide _what_ information to communicate to their neighbors and _how_ to use the received information to take appropriate actions. |
| 81 | +The motivation of designing this library is to study the coverage control problem as a canonical problem for the decentralized navigation of robot swarms. |
| 82 | +We develop the learnable PAC (LPAC) architecture that can learn to process sensor observations, communicate relevant information, and take appropriate actions. |
| 83 | + |
| 84 | +## Architecture |
| 85 | +The learnable Perception-Action-Communication (LPAC) architecture is composed of three different types of neural networks, one for each module of the PAC system. |
| 86 | +1. In the perception module, a convolution neural network (CNN) processes localized IDF observations and generates an abstract representation. |
| 87 | +2. In the communication module, a GNN performs computation on the output of the perception module and the messages received from neighboring robots. |
| 88 | +It generates a fixed-size message to communicate with the neighbors and aggregates the received information to generate a feature vector for the action module of the robot. |
| 89 | +3. In the action module, a shallow multilayer perceptron (MLP) predicts the control actions of the robot based on the feature vector generated by the GNN. |
| 90 | + |
| 91 | +\htmlonly |
| 92 | +<img class="center" style="width: 80%; margin-left: auto; margin-right: auto;" src="learnable_pac.png"/> |
| 93 | +<figcaption>Learnable Perception-Action-Communication (LPAC) architecture: |
| 94 | +The three modules are executed on each robot independently, with the GNN in the communication module facilitating collaboration between robots. |
| 95 | +</figcaption> |
| 96 | +\endhtmlonly |
| 97 | + |
| 98 | +> [LPAC: Learnable Perception-Action-Communication Loops with Applications to Coverage Control.](https://doi.org/10.48550/arXiv.2401.04855) |
| 99 | +> Saurav Agarwal, Ramya Muthukrishnan, Walker Gosrich, Vijay Kumar, and Alejandro Ribeiro. |
| 100 | +> arXiv preprint arXiv:2401.04855 (2024). |
| 101 | +
|
0 commit comments