GAP9Shield Module: Nano-Drone Innovation

• GAP9Shield is a nano-drone module that integrates computation, vision, ranging, and wireless streaming into a compact board for enhanced aerial autonomy.
• It features a 9-core RISC-V SoC with an NE16 AI accelerator achieving 150GOPS, a 5MP camera, and five-direction ToF sensors for robust sensing and processing.
• Designed for low power consumption, the module enables real-time onboard AI, object detection, mapping, and agile navigation in energy-constrained environments.

The GAP9Shield is a nano-drone-compatible module integrating advanced computation, vision, ranging, and wireless connectivity components into a compact and lightweight board. Designed to address the performance and energy limitations typical in nano-drone platforms, GAP9Shield enables on-board AI, real-time object detection, and robust environment sensing within a power envelope suitable for palm-sized aerial vehicles. Its architecture centers on a GAP9 System-on-Chip (SoC) delivering 150GOPS performance and pairs this with a 5MP camera, multi-directional time-of-flight (ToF) sensing, and high-throughput wireless streaming functionalities, facilitating complex tasks such as navigation, mapping, and obstacle avoidance in highly constrained environments (Müller et al., 27 Jun 2024).

1. Hardware Architecture

The GAP9Shield assembles several tightly integrated hardware units to deliver high performance in a miniaturized form factor:

• GAP9 SoC: This System-on-Chip features a 9-core RISC-V compute cluster augmented by an NE16 AI accelerator and transprecision floating-point support (IEEE 32/16, bfloat16). The cluster achieves 15.6 GOPs for DSP and 32.2 GMACs for ML operations at an energy efficiency of 330 μW/GOP. A single-core RISC-V controller orchestrates peripheral I/O, offloading intensive computations to the main cluster. Memory expansion is provided through APS256XXN-OBR PSRAM (volatile) and Macronix MX25UM51245GXDI00 (nonvolatile).
• 5MP OV5647 Camera: This low-voltage CMOS sensor supports QSXGA to QVGA resolutions, outputting 8/10-bit raw RGB via a single MIPI CSI2 interface. While hardware permits up to 120 fps at QVGA, actual rates on GAP9Shield are 15 fps (QVGAs) and 45 fps (VGAs), due to current driver and interface constraints.
• NINA-W102 WiFi-BLE: ESP32-based, dual-core module supporting 2.4 GHz WiFi and Bluetooth. The dual-core system allows WiFi communication to operate concurrently with SPI interfacing to the GAP9, optimizing both telemetry and image streaming throughput.
• 5D VL53L1 ToF Array: Five VL53L1 ToF sensors are deployed for omnidirectional ranging (front, back, left, right, up) with a range up to 400 cm. Multiplexed sampling across I2C attains approximately 40 Hz aggregate.

These components are integrated onto a single board, lowering total system weight and reducing the physical footprint compared to prior sets of separate AI and ranging modules.

2. Performance Metrics

GAP9Shield’s metrics are optimized for AI workloads, rapid sensing, and minimal resource consumption, as detailed below:

Metric	Achieved Value	Comparative Note
Compute Power (SoC)	150 GOPS	Outperforms GAP8, monochrome-only decks
RGB QVGA Streaming Rate	7 fps	+20% vs. grayscale AI-deck
VGA Resolution Streaming	4 fps
WiFi Throughput (TCP/UDP)	8/12 Mbit/s
Total Module Weight	~6 g	20% less than AI+Multi-Ranger decks

• The sample rate improvement for QVGA images represents a 20% gain over prior single-channel, grayscale platforms.
• System weight (approximately 6 g) offers a 20% reduction relative to the AI-deck plus Multi-Ranger deck combination (previously ~7 g), also yielding significant volume savings.
• Transmission speeds via WiFi reach 8 Mbit/s over TCP and 12 Mbit/s over UDP, suitable for high-throughput streaming scenarios.

3. Power Management and Energy Efficiency

Energy constraints are paramount in nano-drone applications. Detailed features ensuring low power operation include:

• Dynamic Frequency Scaling: The GAP9 SoC supports run-time adjustment of clock frequency (up to 370 MHz) and voltage, as well as automatic clock gating.
• Deep Sleep State: Idle power can be reduced to 45 μW.
• Task-Level Consumption: Inference tasks such as YOLO-based object detection can be performed at sub-100 mW, with minimum inference times of 17 ms at ~1.59 mJ. Monte Carlo Localization (MCL) with sensor updates (15 Hz ToF/odometry, 5 Hz camera) averages 23 mW. NanoSLAM mapping operates at approximately 87.9 mW, with sub-250 ms mapping latencies.

The system’s adherence to sub-100 mW envelopes—especially for compute-intensive tasks—is essential for extending flight times and supporting complex real-time mission workloads.

4. Functional Capabilities and Software Applications

GAP9Shield is purpose-built for embedded AI and navigation tasks essential to nano-drone autonomy:

• Object Detection: Optimized YOLO variants deployed on the GAP9 compute cluster achieve detection latencies of 17–38 ms at energy costs in the milliJoule range. Such low-latency detection supports rapid on-board decision-making.
• Localization and Mapping: Algorithms combining semantic cues from object detection and geometric information from the ToF array (e.g., MCL) operate at sensor rates, fusing multi-modal data for robust position estimation. NanoSLAM mapping achieves complete cycles in less than 250 ms, a critical factor for real-time operation in cluttered indoor environments.
• Sensing and Navigation: The five-direction ToF setup provides real-time obstacle data for agile, omnidirectional navigation. The combination of visual and range sensing improves SLAM and obstacle avoidance beyond vision-only or range-only approaches.

These capabilities collectively allow for robust object detection, localization, mapping, and dynamic navigation in compact indoor drones, addressing the historical lack of sufficient on-board computation and sensing (Müller et al., 27 Jun 2024).

5. Comparative Assessment with Existing Nano-Drone Modules

The design and implementation of GAP9Shield are benchmarked against previous nano-drone module configurations:

• Integration: Unlike earlier solutions that coupled separate AI processing and ranging modules (such as those built around the GAP8 with an AI-deck and Multi-Ranger deck), GAP9Shield’s single-board approach unifies high-density AI inference, vision, and multi-directional ranging with wireless streaming.
• Performance Gains: The 150GOPS computation capacity, improved memory resources, and use of a 5MP HD camera (with RGB rather than monochrome output) mark significant advances. The module enables a 20% higher RGB sample rate for image streaming at a 20% reduction in system weight.
• Limitations: There is a noted QVGA frame rate constraint due to current driver implementation, indicating that further software optimization is needed to fully realize the hardware’s performance envelope. For high-frequency communications tasks, WiFi transmission at 250–260 mW can dominate energy consumption, which could impact certain use cases where on-board inference is preferred over external data links.

This holistic design approach, pairing sensory density with AI-centric computing in under 6 g, establishes a new reference point for nano-drone compatibility and autonomy.

6. Implications for Nano-Drone Systems and Research Directions

GAP9Shield demonstrates the feasibility and advantages of integrating advanced AI and high-dimensional sensing directly within nano-scale aerial platforms. The architectural co-location of perception, computation, and communications establishes capabilities that formerly required multiple discrete modules or off-board compute. This suggests accelerating trends toward more autonomous, agile, and power-conscious drone solutions, particularly for indoor robotics, search and rescue, and collaborative swarm operation.

A plausible implication is that expanding embedded AI hardware and more efficient multi-modal sensor integration will enable not only richer autonomy in nano-drones but also facilitate new research in swarm robotics, adaptive mission planning, and distributed SLAM in highly constrained form factors. Challenges in software optimization, especially for vision driver integration and energy–throughput balancing, remain areas requiring further investigation (Müller et al., 27 Jun 2024).