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

TL;DR

This paper surveys autonomous navigation for sub-50g nano-UAVs under stringent SWaP constraints, detailing the physics gap and memory bottlenecks that separate nano-scale flight from traditional robotics. It argues for a holistic hardware-software co-design, emphasizing edge AI, quantization/pruning, and neuromorphic control to operate within a few 10s to a few 100 mW of onboard power. The review covers hardware platforms (MCUs, PULP chips, ASICs), sensing modalities (monocular and sparse ToF, event-based sensing), and an edge-centric autonomy stack spanning learning-based perception, planning, control, and swarm coordination, along with robust simulation and deployment toolchains. Open challenges include Sim-to-Real transfer, perception latency, and energy/perception trade-offs, with a roadmap recommending monolithic System-in-Package solutions, event-driven neuromorphic approaches, and on-device continual learning to enable resilient autonomy in GPS-denied environments.

Abstract

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.

Figures (5)

• Figure 1: The evolution of onboard autonomy for nano-UAVs. Left: Severe Size, Weight, and Power (SWaP) constraints that limit sub-50g platforms. Center: The progression of system-on-chip (SoC) architectures from simple microcontrollers to heterogeneous and neuromorphic processors to address the computational gap. Right: Representative platforms and application domains enabled by these advances.
• Figure 2: The "Edge-AI" Autonomy Stack. A hierarchical view of the hardware-software co-design for nano-UAVs. High-bandwidth sensor data flows from Layer 1 into the perception engine (Layer 3) on the AI accelerator (GAP8), which generates compressed state estimates for the real-time control loop (Layer 4) on the MCU (STM32). Middleware (Layer 2) manages the critical data transport between these heterogeneous cores.
• Figure 3: The Sim-to-Real Training Pipeline. A five-step workflow bridging the gap between simulation and reality. (1) Training begins in high-fidelity simulators (e.g., AirSim). (2) Domain Randomization injects visual noise to induce robustness. (3) Policies are trained using Deep RL or Supervised Learning. (4) Models undergo aggressive compression (Int8 quantization, pruning). (5) The optimized binary is deployed to the nano-UAV's limited L1 memory for real-time inference.
• Figure 4: Swarm Communication Architectures. (A) Centralized paradigms (e.g., Crazyswarm) rely on a single ground station, leading to linear latency growth ($O(N)$) and single points of failure. (B) Decentralized mesh paradigms (e.g., swarm gradient bug algorithm) utilize local communication, achieving constant time complexity ($O(1)$) and enabling scalable, robust swarms in GPS-denied environments.
• Figure 5: System-Level Constraints Governing Nano-UAV Autonomy.(A) Physics of scaling: As the characteristic length $L$ decreases, mass scales with $L^3$ while lifting surface area scales with $L^2$, forcing operation in progressively lower aerodynamic efficiency regimes. (B) Power budget allocation: Propulsion dominates the energy budget ($>$95%), leaving less than 100 mW for onboard sensing, computation, and control (the "5% rule"). (C) Latency mismatch: The fast mechanical dynamics of nano-rotors require control rates on the order of 500 Hz, which is several orders of magnitude faster than the inference latency of vision-based CNNs, motivating hierarchical control architectures.