BioMorphic Intelligence Lab Kick-off Event and Symposium

Distributed Airflow Sensing and Processing on Flapping Wing MAVs

The use of airflow information has proven to be crucial for flyers in nature to avoid predators, to assist in the detection of incoming disturbances and to aid flight control stability. We aim to enrich the lightweight, low-power, and effective low-speed airflow sensing instrumentation inspired by the hair-like structures and mechanoreceptors found in nature. These airflow sensors are then integrated onboard a novel flapping wing MAV platform which is subject to size, weight, power limitations (SWaP), and is easily disturbed by external disturbances. Applications in onboard flow measurement and disturbance detection & rejection are being explored.

Neuromorphic Machine Vision

An event-based camera captures the illumination changes and outputs them as sequences of events, allowing for high temporal resolution and low power consumption.  Many of the good-performing algorithms rely on first converting the events to the frames, and then applying standard computer vision methods such as CNNs. However, these methods neglect the sparse nature of the events, resulting in less efficient algorithms with high latency. Graph neural networks (GNN) enable direct processing of the events thanks to their capability to handle irregularly structured data. Hence, this project focuses on developing computationally efficient algorithms by proposing novel and online spatiotemporal GNNs techniques.

Tactile-Based Control for Aerial Robots

Biological manipulator systems, such as the human hand, showcase superior performance with respect to versatility and dexterity compared to state-of-the-art mechanical manipulation systems. In this work the objective is to develop bio-inspired, compliant limbs that act as sensing and actuation tools for the aerial robot, enabling new navigation and control techniques; and enhancing manipulation capabilities with a focus on fragile and delicate objects.

Efficiency and self-organized criticality in SNNs

Spiking Neural Networks (SNNs) are artificial neural networks that more closely mimic biological neuronal functionalities: by processing visual information with binary, sparse and precisely-timed events (spikes), SNNs can process information faster and more efficiently when compared to traditional ANNs, and are thus ideally suited for processing spatio-temporal event-based information from neuromorphic sensors. However, SNNs are still difficult to train, mainly owing to their complex dynamics of neurons and the non-differentiable nature of spike operations. This project aims to address the issue of efficiency and activity normalization in large-scale SNNs, by making use of the theoretically well established phenomenon of self-organized criticality.

Tactile Perception and Manipulation with Drones

Nowadays, object perception and manipulation with drones are mainly based on visual feedback. However, the working environment of the perception components of a drone is very complicated. Tactile feedback can directly perceive some physical properties of objects (such as hardness, roughness, and temperature) and can therefore improve the perception and manipulation precision in harsh environments such as visual occlusion and insufficient light.

Biomorphic Drones with Multi-modal Locomotion

Multimodality enables robots to overcome limitations of single locomotion modes by executing appropriate transitions between different modalities. Excellently synergized in nature, organisms can move, live, and reproduce in challenging environments man-made robots have only dreamt to thrive in. With biomorphic methodology, this dissertation aims to conceptualise, design, implement, and test an arboreal-aerial robot that can interact with its environment. Adaptive technologies will be utilised to achieve hybrid compliance. In the future, this robotic platform may be applied to search and rescue, and real-world ecological and weather monitoring in unstructured environments.

Object Detection using Neuromorphic Systems in Drones

Event cameras are well-known for high dynamic range, high temporal resolution, and low power consumption. However, we are still facing challenges of applying event data on some computer vision tasks such as object detection and object tracking due to the sparsity of events and the absence of diverse event-based datasets. Currently, I am working on: 1) collecting a novel ego-centric event-based dataset for the purpose of object detection, object segmentation, and action segmentation; 2) explore the data augmentation of events. 

Biomorphic Drones for Ecological Monitoring

Forests and their biodiversity enable human life on Earth. Yet our relationship with nature is broken with forests and biodiversity being lost at alarming rates. It is vital that we learn how to protect and conserve these forests, while being able to sustainably exploit them for our demands. So far, we have not managed to do so, as we lack data from within these ecosystems for holistic understanding. With the development of electronic technology, soft material, vision, control and perception technology, drones have become a low-cost, flexible, and an efficient alternative for forest ecological monitoring. However, multi-rotors have some inherent shortcomings, such as high noise emissions and low endurance. Therefore, in this project I will focus on the  design and control of a novel fixed-wing drone inspired by forest dwellers, addressing challenges in noise emission and endurance.