Nano-Uavs Achieve Autonomous Navigation with Sub-100mw Power and <50g Weight

Researchers are tackling the formidable challenge of enabling autonomous navigation for nano-scale unmanned aerial vehicles, devices limited by incredibly tight Size, Weight, and Power (SWaP) constraints, often weighing less than 50g and operating on under 100mW of power. Mahmud S. Zango and Jianglin Lan, both from the University of Glasgow, alongside their colleagues, present a comprehensive review of the latest sensing, control and algorithmic architectures developed for these miniature robots. This work is significant because it critically assesses the shift from traditional navigation techniques to cutting-edge “Edge AI” approaches, such as quantized deep neural networks and neuromorphic control systems, all within these ultra-low-power budgets. By evaluating both hardware and software co-design, and highlighting remaining obstacles in endurance, obstacle avoidance, and real-world deployment, this survey charts a course towards truly autonomous and agile nano-UAVs capable of operating even without GPS.

By evaluating both hardware and software co-design, and highlighting remaining obstacles in endurance, obstacle avoidance, and real-world deployment, this survey charts a course towards truly autonomous and agile nano-UAVs capable of operating even without GPS.

Nano-UAVs navigate via Edge AI innovation, enabling autonomous

This breakthrough distinguishes nano-UAVs fundamentally from conventional robotic systems, demanding innovative approaches to sensing, computing, and control architectures. This comprehensive survey provides a clear roadmap for addressing these gaps, advocating for hybrid architectures that intelligently fuse lightweight classical control methodologies with data-driven perception systems. Such a fusion will enable fully autonomous, agile nano-UAVs to operate effectively in GPS-denied environments, opening up a wealth of possibilities for previously inaccessible applications. The aerospace domain is undergoing a paradigm shift, moving away from large, robust platforms towards swarms of miniaturized agents, driven by the miniaturization of components and the rise of edge computing .
These “Nano-Fliers” have the potential to transform applications like search-and-rescue in collapsed structures and distributed environmental sensing, but realizing true autonomy at this scale requires overcoming severe SWaP limitations, often restricting onboard computation to a fraction of a few Watts. The work categorizes these nano-scale robots not only by their operational domain but also by the aerodynamic regimes they exploit, identifying two primary categories: Rotary-Wing Nano-UAVs, like the Bitcraze Crazyflie (≈27g), and Bio-Inspired Flappers, such as the DelFly Nimble (28g) and the Harvard RoboBee (80mg). A unifying advantage of these diverse architectures is their ability to operate in highly constrained three-dimensional spaces inaccessible to standard UAVs, but the operational reality is governed by a “Physics Gap” where scaling laws diverge from standard aviation models. Operating at low Reynolds numbers (Re 104), air viscosity dominates, necessitating reliance on unsteady aerodynamics and making vehicles susceptible to disturbances like wind gusts and motor vibrations. This exacerbates the “Sim-to-Real” gap, as simulations often fail to capture the stochastic nonlinearities of nano-scale physics, leading to control policies that perform well in silico but fail in reality. Simultaneously, the “SWaP Gap” limits operational endurance to minutes for rotary-wing vehicles and seconds for flapping-wing prototypes, restricting onboard computational power to approximately 5, 15% of the total, often below 100mW.

Nano-UAV Navigation via Quantised Edge AI enables robust

This research synthesises current sensing, computing, and control architectures tailored for these ultra-low-power systems. Experiments employed dense optical flow algorithms to extract motion cues from visual data, enabling relative pose estimation despite limited computational resources. Simultaneously, scientists developed optimised SLAM algorithms that reduce computational complexity while maintaining accurate 3D map construction, critical for navigation in GPS-denied environments. These algorithms were implemented on custom hardware platforms, including ultra-low-power SoCs, to achieve real-time performance within the stringent SWaP constraints.

Furthermore, the team engineered learning-based flight control policies using reinforcement learning, but recognised the significant “Sim-to-Real” transfer challenges inherent in nano-UAVs. To address this, researchers investigated techniques for domain randomisation and adaptive control, aiming to bridge the gap between simulated and real-world performance. The approach enables the creation of robust control policies that can generalise to unseen environments and disturbances. The study highlights the need for innovative solutions that address these challenges, paving the way for the widespread adoption of nano-UAVs in diverse applications, from search-and-rescue to environmental monitoring, all while operating within the demanding SWaP limitations.

Nano-UAV power limits constrain onboard compute capabilities

Research reveals that typical nano-UAV platforms, such as the Crazyflie 2. x, operate with a total power budget of approximately 7, 10W, with aerodynamic lift consuming 95, 96% of this energy. This leaves a minuscule 5, 15% of the total system power for avionics and onboard intelligence, resulting in a strict compute budget of sub-100mW to a maximum of 200, 300mW for all sensing, processing, and control tasks. Every milliwatt consumed by the navigation engine directly reduces flight time; for example, implementing a visual navigation engine like the PULP-Shield with the DroNet CNN decreased flight time by approximately 22%, from ∼440 seconds to ∼340 seconds, due to a 5g payload increase and electrical consumption. Experiments confirm a critical “Memory Wall” hindering deep learning deployment, as state-of-the-art CNNs require hundreds of megabytes for weights and activations, while nano-UAV microcontrollers operate with only kilobytes of memory, standard MCUs feature roughly 192 kB of SRAM, and advanced SoCs like the GAP8 are limited to 512 kB of L2 memory and 64 kB of L1 scratchpad memory.

A standard MobileNetV1 architecture, considered lightweight, demands memory bandwidths exceeding the capabilities of low-power interfaces unless aggressively optimized, necessitating 8-bit quantization and batch-normalization folding techniques. Measurements show that nano-quadrotors with a mass below 33g and a rotational inertia of approximately 1.4 × 10−5kg m2 exhibit extremely fast dynamics, requiring attitude control loops to operate at 500Hz to maintain stability and state estimators to update at 1kHz. The team measured a significant “latency gap” between computational throughput and flight dynamics, with obstacle avoidance CNNs requiring ∼41 million multiply-accumulate (MAC) operations per frame, resulting in unacceptable latency on a 168MHz Cortex-M4 processor. Heterogeneous accelerators, such as the GAP8, can achieve peak throughputs of ∼10 GMAC/s/W, but even with acceleration, complex CNNs like DroNet operate at only 6, 18 frames per second (FPS), yielding a perceptual latency between 55ms and 160ms.

This necessitates hierarchical control architectures where high-level visual navigation generates setpoints for a lower-level, high-frequency PID controller. Recent work demonstrates that replacing CNNs with Spiking Neural Networks (SNNs) for attitude control allows the perception-actuation loop to operate at 500Hz on a standard microcontroller, eliminating the latency penalty associated with deep learning inference. Furthermore, research indicates that ground effect increases thrust during low-altitude hovering, requiring compensation in the control loop, and downwash from peer drones necessitates safety radii of up to 0.6m to prevent destabilization in swarm configurations. High-frequency vibration, induced by micro-motor speeds up to 2500 rad/s, propagates through the vehicle, adding to the challenges of maintaining stable flight.

Nano-UAVs navigate via Edge AI paradigms, enabling autonomous

The development of sub-50g autonomous nano-UAVs represents a demanding engineering challenge in contemporary robotics. Moving to the nano-scale necessitates a shift to a distinct operational regime, where payload and power directly impact flight endurance and sensing capability. Evidence suggests short-term visual navigation in controlled environments is largely resolved, with compressed perception pipelines and modern flight controllers enabling agile, stable flight within the sub-100mW power envelope. However, incremental optimisation of existing robotics techniques is approaching its limits, and further gains may require fundamentally different computational paradigms. Future research should focus on event-based sensing and neuromorphic or bio-inspired architectures, potentially unlocking a new frontier for nano-UAVs by decoupling autonomy from bulky infrastructure. This could enable pervasive, low-risk operation in confined, GPS-denied, or human-populated environments, benefiting applications like search-and-rescue and precision agriculture with truly low-cost aerial sensing platforms.