NeuRonICS Lab — Demos

ARYABHAT-1: India's First Fully Analog AI Processor

Designed & Built in India — A Fully Analog, Energy-Efficient, Performance-Scalable AI Processor

Key Features

  • 8-Core Fully Analog Architecture
  • Technology & Energy Scalable Design
  • Custom Compiler for AI/ML Algorithms
  • Python-based Automated Testing Framework

Technological Highlights

  • Built from transistor to system level
  • User-configurable energy and performance modes
  • Ultra-low power consumption for edge AI tasks
  • Scalable across process technologies

Bird Hotspots: Low-Power Acoustic Classification for Ecological Monitoring

Identifying Bird Habitats using TinyML & Neuromorphic Audio Processing

Key Features

  • TinyML-powered low-power acoustic classification
  • Neuromorphic audio front-end using in-filter computation
  • Template-based SVM model requiring less data and robust to noise
  • Easy deployment on low-power microcontrollers

Technological Highlights

  • TI CC1352 Launchpad (ARM Cortex M4F)
  • MEMS microphone for audio capture
  • Real-time audio buffering and classification
  • On-device bird detection logs

mmWave Radar for Activity Recognition and Localization

Real-time Multi-person Tracking for Healthcare and Surveillance

Key Features

  • Radar-based Localization & Activity Recognition
  • Simultaneous Multi-person Tracking
  • Three-layer CNN for Real-time Classification
  • Python-based GUI for Visualization & Logging

Technological Highlights

  • IWR6843 mmWave radar (60–64 GHz)
  • Robust in low light, dark, or foggy environments
  • Privacy-preserving — uses point cloud data, no images
  • Tracks range, velocity, acceleration & zone-based localization
  • Classifies walking, sitting, resting, jumping, and more

Neuromorphic Ising Machine for NP-Hard Problems

A multi-institute collaboration from the Bangalore Neuromorphic Engineering Workshop (BNEW)

Key Features

  • Quantum-inspired dynamics where noise aids solution exploration
  • Native support for higher-order Ising interactions
  • Highly scalable and hardware-efficient neuromorphic architecture
  • State-of-the-art performance on benchmark problems

Technology Highlights

  • Fully implemented on the RFSoC 4x2 Evaluation kit
  • Fowler-Nordheim annealing provides asymptotic convergence guarantee
  • Quantum-inspired algorithm on CMOS — scalable, high-speed, room-temperature operation
  • Distributed, parallel, and event-driven neuromorphic computing architecture

Inspired by the way physical systems naturally evolve toward minimum-energy states, this neuromorphic Ising machine efficiently searches for optimal solutions to combinatorially hard problems — enabling fast scheduling, logistics, drug discovery, and other large-scale combinatorial optimisation tasks.

HOMI: Ultra-Fast EdgeAI Platform for Event Cameras

High-speed, low-latency AI acceleration for sparse event-based vision

Key Features

  • Over 1000 FPS processing capability for HDR images
  • Novel pre-processing block for generating frames from sparse events while keeping the dynamic range intact
  • Energy-efficient, low-latency AI accelerator that leverages the inherent sparse nature of event data

Applications

  • Navigation for Autonomous Robots
  • High-Speed Drone/Missile Detection
  • Space Situational Awareness (SSA) Applications

Neuromorphic Radar for Gesture Recognition

Bio-inspired, event-driven radar for always-on, low-power gesture recognition

Key Features

  • Event-driven radar architecture that generates spikes only on meaningful motion — no wasteful processing of idle/background signals
  • Bio-inspired asynchronous sigma-delta encoding that converts IF signals into sparse, spike-based representations mimicking retinal and auditory neurons
  • Interrupt-driven processing pipeline that keeps the microcontroller in a low-power state between events
  • Recognizes five distinct gestures: Push-Pull, Slow Wave, Fast Wave, Up-Down, and No Activity

Technology Highlights

  • 24 GHz pulse-Doppler radar paired with a custom in-house neuromorphic sampler board
  • Lightweight neural network deployed on an ARM Cortex-M0 microcontroller with a memory footprint of only ~4 KB
  • Real-time inference directly on spike polarity and timestamps — no ADC, FFT, or spectrogram reconstruction required
  • ≥85% real-time inference accuracy across five gestures collected from 7 users
  • First demonstrated bio-inspired asynchronous sigma-delta encoding framework for radar-based HGR

By replacing continuous ADC sampling and dense spectrogram processing with sparse, event-driven sensing and computation, this neuromorphic radar framework delivers significant savings in power, memory, and latency — making it ideally suited for always-on gesture recognition in resource-constrained embedded and IoT platforms.

Asynchronous High-Speed Tracking of Astronomical Objects using Neuromorphic Cameras

Real-time, frame-free, multi-target clustering and tracking for Space Situational Awareness

Key Features

  • Fully asynchronous, event-driven clustering and tracking operating directly on raw events (no frame binning)
  • Constant-Turn Extended Kalman Filter (CT-EKF) with constant-neighbour search for near-constant per-event complexity
  • Multi-target tracking with inherent suppression of background noise and hot pixels
  • High dynamic range (≥120 dB) operation, resistant to bright-source glare and motion blur

Technology Highlights

  • Deployed on Raspberry Pi 4 with a Prophesee Gen4 (IMX636) event sensor for on-device real-time inference
  • Tracks blobs moving at > 17,000 px/s with per-event EKF latency < 0.7 µs
  • End-to-end clustering + EKF pipeline at ~1 µs/event, < 12 px localization error, F1-score > 0.9
  • Validated via LED-matrix multi-target simulations and 8-inch telescope trials (Starlink passes, drifting star fields)

By leveraging the µs-scale temporal resolution and sparse output of neuromorphic sensors, this edge-computing platform enables low-latency, low-data-rate tracking for autonomous space situational awareness on both ground- and space-based platforms.

SUSHRUT — A Reconfigurable, AI-Enabled Handheld Ultrasound Device for Super-Resolution Imaging

Portable, palmtop ultrasound with hardware-accelerated AI for high-resolution diagnostic imaging

Key Features

  • Novel non-linear hardware beamforming algorithms offering superior resolution and contrast, innovated for handheld systems
  • Real-time adaptability to different transducers
  • 3D imaging
  • Raw data extraction
  • Handheld 3D sweep ultrasound

Technology Highlights

  • Fast HW AI accelerator
  • 128 TX/RX channels
  • Embedded MCU for video generation and image processing
  • Low power consumption < 6 W
  • Compact 5.9 × 2.3 inch handheld/palmtop form factor

Explore More Demonstrations

Watch more videos and project showcases from the NeuRonICS Lab on our YouTube channel.

Visit Our YouTube Channel →