Bitcraze Crazyflie Nano-Drone

Updated 19 January 2026

Crazyflie Nano-Drone is a modular, open-source quadrotor weighing under 30 g, featuring plug-and-play decks for sensing, computation, and control.
It offers comprehensive onboard sensing and perception with IMU, optical flow, ToF sensors, and camera-based localization to ensure robust, real-time navigation.
Its integrated AI pipelines and advanced control strategies enable aggressive maneuvers, swarm coordination, and autonomous decision-making in resource-constrained environments.

The Bitcraze Crazyflie nano-drone is a modular, open-source, sub-30 g quadrotor platform that has become the reference standard for academic and industrial research on resource-constrained aerial robotics, swarm control, onboard deep learning, and robust autonomous navigation in highly constrained environments. Its architecture, expansion capabilities, and support for advanced AI-centric workflows across sensing, planning, and control have enabled experimental research in large-scale multi-drone systems, onboard AI-based autonomy, sim-to-real transfer for visual pipelines, and aggressive closed-loop control.

1. Hardware Architecture and Modularity

The Crazyflie platform, in its prevalent 2.x/2.1 revision, consists of a 27–30 g micro-quadcopter with a 10 × 10 cm airframe, STM32F405 main microcontroller (168 MHz, 192 kB SRAM, 1 MB Flash), nRF51 or nRF51822 radio MCU for a 2.4 GHz radio link, and provisions for up to two modular “decks.” The decks are mezzanines that expand capabilities via SPI, I²C, UART, and direct GPIO, supporting “plug-and-play” sensors and co-processors:

Sensor decks: Flow deck (optical flow + ToF), Lighthouse deck (Vr-based mm-precision localization), Multiranger (VL53 ToF quad array), and custom sensor boards (e.g., 8×8 ToF arrays for depth mapping (Crupi et al., 2023)).
Compute decks: AI-deck, embedding a GreenWaves GAP8 SoC (RISC-V, 8+1 cores, 512 kB L2, 64 kB L1) for parallel low-power inference; ESP32 radio for Wi-Fi integration.
Power and endurance: Each vehicle is powered by a 240–300 mAh LiPo cell, yielding 4–7 min flight with standard payloads (“with AI-deck: 4.4 g”; “with sensor decks, up to 34 g fully loaded”) (Bouwmeester et al., 2022, Kazim et al., 2023). Modular deck stacking allows customized tradeoffs between perception, compute, and flight time.

Mechanical integration is achieved entirely via snap-in carbon-fiber frame plates and M2 standoffs, enabling rapid hardware extension/replacement.

2. Onboard Sensing and Perception Capabilities

Out-of-the-box, the Crazyflie supports 6-DOF IMU (accelerometers/gyroscopes), optical flow and ToF-based barometric altitude estimation. Expansion decks enable:

Monocular vision: HiMax HM01B0 QVGA (320×240 or 160×160 px) 60 Hz grayscale camera, connected via parallel CPI to the GAP8 or onboard MCUs (Bouwmeester et al., 2022, Palossi et al., 2021, Crupi et al., 2023).
Multi-zone depth: 8×8 ToF (ST VL53LC5CX, 313 mW, 15 Hz), providing 64-range measurements in 45° FOV (Crupi et al., 2023).
Relative localization: Visual fiducial-based or dense FCNN-based direct pose estimation to track other Crazyflie units at up to 39 Hz, with median (x,y,z) tracking error under 10 cm for battery duration, using onboard quantized CNNs (Crupi et al., 2024).
Human following/pose estimation: CNNs trained for either direct regression or fusion (depth+vision), supporting closed-loop human-aware navigation at onboard rates exceeding 45–135 Hz within a <100 mW compute budget (Palossi et al., 2021, Crupi et al., 2023, Cereda et al., 12 Jan 2026).

The sensory stack, including convolutional dense optical flow estimation (Bouwmeester et al., 2022), enables vision-based navigation and real-time environmental interaction within the strict energy, size, and weight constraints inherent to gram-scale aerial robots.

3. Onboard and Distributed AI Pipelines

Support for ultra-low-power DNN inference is provided by the custom AI-deck with GAP8 SoC. Key onboard AI workflows:

DNN visual navigation: End-to-end CNN-based pipelines for collision avoidance, obstacle detection, relative pose estimation, and racing, realized using quantized (e.g., 8-bit) models on GAP8 (Palossi et al., 2018, Palossi et al., 2019, Palossi et al., 2021, Bouwmeester et al., 2022, Crupi et al., 2024).
Distributed computing: Edge-offload paradigms, where high-capacity DNNs (e.g., SSD-MobileNetV2 for object detection) run on external workstations, with only planning/control on the drone; supports closed-loop rates of 5–8 Hz with ~170 ms total latency in moderately cluttered environments at 1 m/s (Sartori et al., 8 May 2025).
Coroutine-based real-time pipelines: NanoCockpit framework enables time-optimal pipelining across camera acquisition, multi-core inference, DMA-based memory transfers, and Wi-Fi streaming, achieving zero-overhead frame/actuation latency and maximizing closed-loop control frequencies (Cereda et al., 12 Jan 2026).
Reinforcement learning: Deep Q-learning for obstacle-laden source-seeking executed entirely on the STM32 Cortex-M4 (100 Hz, ≈620 parameter float DQN), with only 0.14 W compute overhead, establishing that non-vision networks can run with extremely minimal power and memory on the native microcontroller alone (Duisterhof et al., 2019).

Quantization-aware training, memory-tiling (DORY/AutoTiler), and low-level SIMD kernel optimization are standard for all vision-centric CNN deployments, with measured end-to-end inference times as low as 7 mJ/frame at 6 Hz (DroNet) and throughput up to 135 fps at 86 mW (PULP-Frontnet) (Palossi et al., 2018, Palossi et al., 2021).

4. Advanced Control and Planning Methodologies

The Crazyflie has enabled agile control research ranging from simple PID loops to geometric and NMPC controllers:

Aggressive trajectory tracking: Embedded nonlinear MPC (acados-generated, onboard GAP8), realizing full quaternion rigid-body dynamics, direct multiple shooting, horizon N=10, solved at 10 Hz for onboard aggressive maneuvers (e.g., helical climbs) with centimeter-level accuracy (Kazim et al., 2023).
Gaussian Process–augmented geometric control: For aggressive backflip maneuvers, both feedforward and robust adaptive feedback controllers exploit GP models for disturbance estimation and trajectory parameterization, validated in real flight (10/10 flips, final error <0.04 m with disturbance payloads) (Antal et al., 2022).
Collision-free multi-agent formation: Distributed outer-loop control laws, combining finite cut-off potential functions for inter-agent and obstacle avoidance with Laplacian-based consensus tracking, achieve provably stable and collision-free time-varying formations at up to 0.4 m/s (minimum 0.4 m pairwise distance) (Nguyen et al., 2021).

The modular control architecture allows both hard real-time (1 kHz) loops for attitude stability and distributed, software-upgradable higher-level planners.

5. Communication, Swarming, and Integration

The Crazyflie natively supports both proprietary (Crazyradio PA, ESB protocol, 2–4 ms median RTT) and open (Wi-Fi via ESP32, Pi Zero W, IP/UDP/PRRT, median RTT ~9 ms with optimized bridges) remote control links (Böhmer et al., 2020). Integration with open middleware and predictable low-latency stacks facilitates:

Swarming research: Large-scale (>10 node) multi-agent experiments, supporting both relative localization (onboard CV) and formation control, with automated assignment of setpoints via ROS/Matlab/Simulink, and flexibility for heterogeneous networks (Nguyen et al., 2021, Shang et al., 2019).
Networked autonomy: Streaming control and perception data over Wi-Fi and mesh IP networks for distributed mapping, exploration, or real-time external processing, fully leveraging standard internet protocols (Böhmer et al., 2020, Sartori et al., 8 May 2025).

Multiplexed communication with external computers, as well as inter-drone visual/camera-based relative pose estimation, supports dynamic distributed tasks (e.g., drone racing, formation flying, spacecraft-proxy maneuvers (Barcena et al., 2024)) with robust, reproducible reactive behavior.

6. Experimental Methodologies and Performance Benchmarks

The platform’s well-documented experimental protocols facilitate direct reproducibility:

Metric	Value/range	Reference
Baseline airframe mass	27–30 g	(Antal et al., 2022)
Onboard compute budget	≤100 mW (GAP8); <10 mW (STM32 only tasks)	(Palossi et al., 2021)
Max AI inference throughput	135 fps (PULP-Frontnet); 18 fps (DroNet)	(Palossi et al., 2021, Palossi et al., 2018)
Optical flow CNN (NanoFlowNet)	5.57–9.34 fps, 7–9 M MACs, 170k params	(Bouwmeester et al., 2022)
NanoDrone-to-Drone Relocalization (FCNN)	39 Hz, 101 mW, median 3D err <0.1 m	(Crupi et al., 2024)
Ultra-low cost visual navigation	6 Hz @ 64 mW, up to 18 Hz @ 284 mW	(Palossi et al., 2018, Palossi et al., 2019)
Communication RTT (proprietary/open)	4 ms / 9 ms (median, optimized)	(Böhmer et al., 2020)
Aggressive backflip, max error (robust)	<0.04 m drift (10/10 flips)	(Antal et al., 2022)
Real-time planning latency (split DNN)	~170 ms end-to-end @ 8 FPS	(Sartori et al., 8 May 2025)

Field experiments are extensively documented across scenarios (collision avoidance, cluttered maze, human following, drone-to-drone pursuit, spacecraft formation emulation, and random waypoint inspection), and support ground-truth tracking via Vicon/OptiTrack/Loco Positioning at up to 200 Hz (Cereda et al., 12 Jan 2026).

7. Research Impact and Future Directions

The Bitcraze Crazyflie ecosystem underpins a large cross-section of nano-aerial vehicle research in perception, control, AI-acceleration, and swarm robotics. All core platforms, expansion decks, firmware, and major AI pipelines (DORY, PULP-NN, NanoCockpit, PULP-Frontnet, FCNN, NanoFlowNet, etc.) are released open-source, with reproducibility guidelines, exact hardware BOMs, codebases, and pre-trained models available for the research community (Palossi et al., 2018, Cereda et al., 12 Jan 2026, Crupi et al., 2023). This combination of hardware modularity, accessible real-time firmware, and tailored AI-centric toolchains accelerates the design, deployment, and benchmarking of advanced algorithms for fully onboard autonomy, even in domains such as sim-to-real pose estimation, multi-modal sensor fusion, and robust model-predictive control under severe computation constraints.

Emerging directions include: increasing edge AI-deck integration density, exploiting deeper co-pipelining for multi-task inference, integrating more heterogeneous sensory modalities, closed-loop sim-to-real visual learning, distributed collaborative mapping, and hybrid onboard–networked autonomy for massive drone swarms (Cereda et al., 12 Jan 2026, Sartori et al., 8 May 2025, Crupi et al., 2023).

References: (Palossi et al., 2018, Shang et al., 2019, Palossi et al., 2019, Duisterhof et al., 2019, Böhmer et al., 2020, Palossi et al., 2021, Nguyen et al., 2021, Bouwmeester et al., 2022, Antal et al., 2022, Crupi et al., 2023, Kazim et al., 2023, Crupi et al., 2024, Barcena et al., 2024, Sartori et al., 8 May 2025, Cereda et al., 12 Jan 2026)