Tinynav

Abstract

Achieving real-time, autonomous obstacle avoidance and path planning on low-cost robotics platforms, such as those based on microcontrollers like Arduino, is a significant challenge due to strict computational and memory constraints. Existing approaches often rely on manual control, predefined algorithms, or more powerful processors like single-board computers, which limits accessibility and scalability. This project addresses this gap by leveraging recent advances in microcontroller hardware and TinyML frameworks to enable on-device neural network inference for autonomous navigation.We evaluate the feasibility of deploying a lightweight Convolutional Neural Network (CNN) augmented with temporal context on a low-resource microcontroller using Time-of-Flight (ToF) sensor data. Data are collected in a controlled circuit composed of simple walls and box-shaped obstacles, with the robot manually driven to capture several hours of sensor readings, reducing variability and ensuring reliable generalization. The architecture combines a 2D CNN for spatial feature extraction with a Long Short-Term Memory (LSTM) layer to capture temporal dependencies across sequential frames. CNN features are extracted from consecutive frames using a TimeDistributed approach, after which the LSTM predicts steering and acceleration commands. Model compression techniques, including pruning and quantization, are applied to meet flash memory and real-time inference constraints.Our results demonstrate that running neural networks for autonomous navigation on low-cost microcontrollers is practical, even under limited computational resources. This highlights the potential of TinyML to enable intelligent behavior on affordable, low-power devices. By showing that meaningful autonomy can be achieved without high-end hardware, this work expands the applicability of TinyML in real-world robotic systems and opens the door to more accessible and adaptive low-cost robotics solutions.