NaviSlim: Adaptive Context-Aware Navigation and Sensing via Dynamic Slimmable Networks

Tim Johnsen,Marco Levorato

2024-05-16

Abstract:Small-scale autonomous airborne vehicles, such as micro-drones, are expected to be a central component of a broad spectrum of applications ranging from exploration to surveillance and delivery. This class of vehicles is characterized by severe constraints in computing power and energy reservoir, which impairs their ability to support the complex state-of-the-art neural models needed for autonomous operations. The main contribution of this paper is a new class of neural navigation models -- NaviSlim -- capable of adapting the amount of resources spent on computing and sensing in response to the current context (i.e., difficulty of the environment, current trajectory, and navigation goals). Specifically, NaviSlim is designed as a gated slimmable neural network architecture that, different from existing slimmable networks, can dynamically select a slimming factor to autonomously scale model complexity, which consequently optimizes execution time and energy consumption. Moreover, different from existing sensor fusion approaches, NaviSlim can dynamically select power levels of onboard sensors to autonomously reduce power and time spent during sensor acquisition, without the need to switch between different neural networks. By means of extensive training and testing on the robust simulation environment Microsoft AirSim, we evaluate our NaviSlim models on scenarios with varying difficulty and a test set that showed a dynamic reduced model complexity on average between 57-92%, and between 61-80% sensor utilization, as compared to static neural networks designed to match computing and sensing of that required by the most difficult scenario.

Robotics,Artificial Intelligence,Machine Learning

What problem does this paper attempt to address?

The paper attempts to address the problem of achieving efficient and adaptive navigation and perception capabilities in resource-constrained small autonomous flying vehicles, such as micro drones. Specifically, micro drones face significant limitations in computational power and energy reserves, making it challenging to support complex modern neural network models that are crucial for autonomous operation. The main contribution of the paper is the proposal of a new neural navigation model called NaviSlim, which can dynamically adjust the usage of computational and perception resources based on the current environment's difficulty, the current trajectory, and the navigation goals.

The design of NaviSlim is based on a gated variable-width neural network architecture. Unlike existing variable-width networks, it can dynamically select a "slimming factor" to autonomously adjust the model's complexity, thereby optimizing execution time and energy consumption. Additionally, compared to existing sensor fusion methods, NaviSlim can dynamically select the power levels of onboard sensors without switching between different neural networks, autonomously reducing the time and energy consumption during the sensor acquisition process.

Through extensive training and testing in the Microsoft AirSim simulation environment, the researchers evaluated the performance of the NaviSlim model in scenarios of varying difficulty. The results showed that, compared to static neural networks, the NaviSlim model reduced model complexity by an average of 57-92% and sensor utilization by 61-80%. This indicates that NaviSlim can adaptively adjust the resources used (including power, time, and energy) based on the perceived environmental context, significantly reducing resource consumption while maintaining high navigation accuracy.