ColibriUAV: An Ultra-Fast, Energy-Efficient Neuromorphic Edge Processing UAV-Platform with Event-Based and Frame-Based Cameras

Abstract

ColibriUAV is an ultra-fast, energy-efficient UAV platform that combines traditional frame-based cameras with dynamic vision sensors (DVS) for robust, low-latency perception. The system is centered around Kraken, a novel low-power RISC-V System on Chip (SoC) featuring dual hardware accelerators optimized for Spiking and Deep Ternary Neural Networks. This integration achieves state-of-the-art efficiency, processing event data at 7200 frames per second with a subsystem power consumption of only 10.7 mW, making it suitable for low-latency autonomous nano-drones.

Report

Key Highlights

• Hybrid Sensing Platform: ColibriUAV integrates both frame-based (RGB) and dynamic event-based (DVS) cameras to ensure low-latency, robust perception for Unmanned Aerial Vehicles (UAVs).
• RISC-V Core: The platform is powered by the novel Kraken SoC, a low-power RISC-V chip specifically designed for near-sensor neuromorphic edge processing.
• High Efficiency: The neuromorphic event data subsystem achieves a throughput of 7200 frames of events per second while consuming only 10.7 mW.
• Performance Gain: This specialized approach is reported to be over 6.6 times faster and a hundred times less power-consuming than common data reading methods (like USB interface).
• Total Power Budget: The overall sensing and processing power consumption is maintained below 50 mW, resulting in latency in the milliseconds range, ideal for nano-drones.

Technical Details

• Platform Name: ColibriUAV.
• Main Processor: Kraken, a novel low-power RISC-V System on Chip (SoC).
• Accelerators: Kraken includes two dedicated hardware accelerators targeting:
  1. Spiking Neural Networks (SNNs).
  2. Deep Ternary Neural Networks (TNNs).
• DVS Integration: Kraken features an integrated, dedicated hardware interface optimized for direct communication with the DVS camera, significantly improving energy efficiency and reducing latency compared to standard bus interfaces.
• Energy Metrics (Event Processing): 10.7 mW power consumption for event data reading.
• Latency Goal: Achieving millisecond-range latency for end-to-end processing.

Implications

• Validation of RISC-V in Edge AI: The Kraken SoC demonstrates the successful deployment of customized RISC-V architectures specifically tailored for complex, real-time edge workloads like neuromorphic vision. This strengthens RISC-V's competitive position against established instruction sets in power-constrained AI applications.
• Accelerating Neuromorphic Computing: By tightly integrating specialized hardware accelerators for SNNs and TNNs onto the RISC-V SoC, the platform validates advanced, low-power AI techniques that are crucial for replicating bio-inspired intelligence at the edge.
• Setting New Efficiency Benchmarks: The colossal power saving (100 times less) and speed increase (6.6 times faster) achieved by the dedicated DVS interface proves the value of custom hardware/IP built around the RISC-V core for eliminating I/O bottlenecks common in conventional systems.
• Enabling Nano-Drones Autonomy: The extremely low power consumption (sub-50 mW total processing) and millisecond latency make sophisticated autonomous navigation and perception viable for smaller, lighter, and longer-flying UAVs (nano-drones) previously constrained by computational overhead.