Paper-Conference |

LiMPNet: Lightweight Multi-sensor Perception and DRL Navigation for Tiny Drones in Mapless Environments

Sat, 01 Nov 2025 00:00:00 +0000

Abstract

Autonomous tiny drones are severely constrained by size, weight, power, and onboard computation, making reliable autonomous navigation particularly challenging. This work presents LiMPNet, a lightweight navigation framework that combines monocular vision, laser range sensing, and deep reinforcement learning to enable safe mapless navigation in cluttered environments. The system integrates a YOLOv8n-based obstacle detector, multi-ranger distance sensing, IMU-based state estimation, and a PPO navigation policy within a ROS and Gazebo simulation framework. Experimental results demonstrate reliable autonomous obstacle avoidance using only lightweight sensors, achieving a 100% success rate (112/112) in a simple environment and 35% success (7/20) in a densely cluttered environment. These results demonstrate that efficient autonomous navigation is feasible on highly resource-constrained aerial robots while maintaining low computational complexity.

Lightweight Temporal Consistency for Grid-Based Obstacle Detection in Edge Devices

Mon, 01 Sep 2025 00:00:00 +0000

Abstract

We present a lightweight obstacle detection framework designed for deployment on resource-constrained edge devices, ranging from desktop GPUs to Raspberry Pi and microcontrollers. Instead of relying on conventional object detection outputs, our method converts images into a compact 6 × 8 binary obstacle matrix that enables efficient spatial reasoning for autonomous systems. To improve robustness against motion blur, occlusion, and lighting variations, we introduce a lightweight temporal smoothing mechanism that significantly enhances detection consistency while adding virtually no computational overhead. The framework is evaluated using YOLOv8n, SSD-MobileNet v2, and FOMO across multiple hardware platforms, demonstrating up to a 13.38 percentage point improvement in detection accuracy, while matrix projection and smoothing together require less than 0.1 ms of additional processing time. The proposed approach provides an efficient perception module for edge AI, TinyML, autonomous robots, and micro-drones.

SunSift: Solar-Powered Intelligent Sensing through Informative Sample Selection

Mon, 01 Sep 2025 00:00:00 +0000

Abstract

SunSift is a lightweight framework for batteryless CNN inference on energy-harvesting IoT devices. It uses intelligent voltage-based state transitions and selective MRAM checkpointing to support intermittent execution under frequent power failures. The system is implemented using Arduino Nano 33 BLE boards and TensorFlow Lite for Microcontrollers, achieving 71.51% accuracy in identifying informative samples from CIFAR-10 for distributed federated learning. This work demonstrates a practical path toward sustainable, batteryless edge intelligence.

Autonomous Integrated Sensing and Processing for BioRadar: Advancing Non-Invasive Biomedical Monitoring and Signal Analysis

Wed, 01 Jan 2025 00:00:00 +0000

This work presents an integrated sensing and processing system for continuous-wave BioRadar, enabling non-invasive vital sign monitoring through embedded signal processing.

Soil Excavation Analysis for Autonomous Tunnel Burrowing Robots Using an Instrumented Robotic Manipulator

Wed, 01 Jan 2025 00:00:00 +0000

This work investigates autonomous soil excavation for tunnel-burrowing robots using an instrumented robotic manipulator. The study analyzes excavation forces, soil interaction, and robotic manipulation strategies to improve autonomous underground construction.