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

TL;DR

NaviSlim addresses autonomous navigation for micro-drones under strict compute and energy limits by introducing a context-aware framework that dynamically scales both neural computation and onboard sensing. The system comprises NaviSlim-C, a computing-adaptive module, and NaviSlim-S, a sensing-adaptive module, built on universally slimmable networks and guided by knowledge distillation, reinforcement learning, and curriculum strategies. Through training and evaluation in Microsoft AirSim, NaviSlim achieves substantial reductions in model complexity (57-92%) and sensor usage (61-80%) while preserving length-optimal navigation, demonstrating adaptive resource allocation aligned with environmental difficulty. This approach enables faster reaction times and better energy efficiency for resource-constrained aerial autonomy, with broad implications for real-time adaptive perception and control in small UAVs.

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.

Figures (17)

• Figure 1: High-level schematics of the considered sensing-computing-control pipeline. The neural network processes the input of forward and downward facing depth sensors and GPS to produce motion commands that fly the drone on an ideally length-optimal path from point A to B.
• Figure 2: High-level schematics of the considered sensing-computing-control pipeline, now with an introduced attention mechanism and the ability to predict context-aware computing operations and sensing commands.
• Figure 3: Two maps from Microsoft AirSim: on the left is "Blocks" which contains static objects with arbitrary shapes and sizes, and on the right is "City" which contains both static and dynamic objects expected to be encountered in the real world.
• Figure 4: NaviSlim: our novel solution for a context-aware framework capable of adapting resource allocation to that which is required by the difficulty of the current scenario. Shown is our specific implementation. The shapes with dotted lines represent components capable of adaptable resource allocation.
• Figure 5: Procedure used to "slim" a universally slimmable neural network. The variable $\rho$ is used to scale the number of active nodes in each hidden layer. In this example, there are 4 nodes in the first hidden layer and 2 in the second -- this is the super-network. Also in this example, we use $\rho=0.3$ so that 2 nodes are deactivated in the first layer and 1 in the second -- this is a sub-network. Note that all weights connected to the deactivated nodes, represented by the lines in between the hidden layers, are also severed.