Autonomous Navigation at the Nano-Scale: Algorithms, Architectures, and Constraints

Autonomous navigation for nano-scale unmanned aerial vehicles (nano-UAVs) is governed by extreme Size, Weight, and Power (SWaP) constraints (with the weight < 50 g and sub-100 mW onboard processor), distinguishing it fundamentally from standard robotic paradigms. This review synthesizes the state-of-the-art in sensing, computing, and control architectures designed specifically for these sub- 100mW computational envelopes. We critically analyse the transition from classical geometry-based methods to emerging “Edge AI” paradigms, including quantized deep neural networks deployed on ultra-low-power System-on-Chips (SoCs) and neuromorphic event-based control. Beyond algorithms, we evaluate the hardware-software co-design requisite for autonomy, covering advancements in dense optical flow, optimized Simultaneous Localization and Mapping (SLAM), and learning-based flight control. While significant progress has been observed in visual navigation and relative pose estimation, our analysis reveals persistent gaps in long-term endurance, robust obstacle avoidance in dynamic environments, and the “Sim-to-Real” transfer of reinforcement learning policies. This survey provides a roadmap for bridging these gaps, advocating for hybrid architectures that fuse lightweight classical control with data-driven perception to enable fully autonomous, agile nano-UAVs in GPS-denied environments.

💡 Research Summary

This review paper provides a comprehensive synthesis of the state‑of‑the‑art in autonomous navigation for nano‑scale unmanned aerial vehicles (nano‑UAVs), defined as platforms weighing less than 50 g and operating on less than 100 mW of onboard power. The authors begin by outlining the dual challenges that set nano‑UAVs apart from conventional UAVs: a “Physics Gap” arising from low Reynolds‑number aerodynamics, which makes these vehicles highly susceptible to viscous effects, gusts, and motor‑induced vibrations, and a “SWaP Gap” that limits the power budget available for computation to roughly 5‑15 % of total consumption.

The paper categorises nano‑UAV platforms into rotary‑wing (e.g., Bitcraze Crazyflie, Black Hornet Nano) and bio‑inspired flappers (e.g., RoboBee, DelFly). It then surveys the hardware landscape, moving from classic microcontroller units (STM32F4, Cortex‑M7) to parallel ultra‑low‑power (PULP) architectures such as GAP8 and GAP9, and finally to application‑specific integrated circuits (ASICs) like Navion and the heterogeneous Shaheen SoC. Table 1 quantifies each substrate’s peak throughput, energy efficiency, memory footprint, and typical power draw, demonstrating that modern RISC‑V clusters can deliver 20‑300 GOPS while staying under 200 mW, and that ASIC accelerators can perform visual‑inertial odometry at 171 fps for only 24 mW.

On the algorithmic side, the authors dissect the “Onboard Autonomy Stack”. Perception is moving from dense optical‑flow and frame‑based cameras to event‑based vision, quantised convolutional neural networks (TinyML), and spiking neural networks (SNNs). They cite concrete examples: a GAP8‑based AI‑deck running the Frontnet CNN at 135 fps for 86 mW, and a Cortex‑M7 implementation of an SNN attitude controller achieving 500 Hz loops with negligible power. State estimation combines lightweight EKF‑VIO pipelines with inertial data, while control ranges from classic PID/LQR to neuromorphic SNN controllers. Planning and decision‑making employ lightweight graph‑search (RRT*, A*) and reinforcement‑learning (RL) policies, yet the authors stress the persistent “Sim‑to‑Real” gap that hampers transfer of RL policies trained in high‑fidelity simulators to physical nano‑UAVs.

The review highlights that the severe SWaP constraints preclude the use of high‑resolution LiDAR or GPU‑based VSLAM; instead, navigation must rely on noisy, low‑resolution Time‑of‑Flight or optical‑flow data. Battery capacities (<250 mAh) limit flight time to minutes for rotary‑wing platforms and seconds for flappers, reinforcing the need for ultra‑efficient computation. The authors also discuss swarm intelligence as a means to distribute computational load and achieve complex mission objectives through local interaction rules.

Identified research gaps include (1) long‑duration endurance under sub‑100 mW budgets, (2) robust obstacle avoidance in dynamic, cluttered environments using only sparse visual cues, (3) reliable Sim‑to‑Real transfer of RL policies, and (4) scalable deployment of ASIC‑level VIO accelerators. To address these, the paper proposes a roadmap centred on hybrid architectures that fuse lightweight classical control with data‑driven perception: (i) co‑design of event‑based sensors and SNN accelerators, (ii) aggressive model compression (quantisation, pruning, knowledge distillation) for multi‑task TinyML models, (iii) domain‑adaptation techniques integrated into RL pipelines, (iv) broader adoption of low‑power VIO ASICs, and (v) development of decentralized swarm control frameworks.

In conclusion, the authors argue that bridging the physics, computational, and communication constraints through such hybrid, bio‑inspired, and neuromorphic solutions will enable fully autonomous, agile nano‑UAVs capable of operating in GPS‑denied, cluttered environments, opening new application domains ranging from disaster response to precision agriculture.