Adaptive and Efficient Robot Positioning

Project dates: 2024 - Ongoing
Introduction
This project is an Australian Research Council Discovery Early Career Researcher Award (DECRA) Fellowship, awarded to Dr Tobias Fischer in 2023. DECRA Fellows are “outstanding early-career researchers with demonstrated capacity for high-quality research and emerging capability for leadership and supervision”. The DECRA supports the Fellow and a team of PhD students for 3 years. In this case, the ARC contributes a $462,000 fellowship, with another $300,000 cash funding from QUT, supporting two PhD student scholarships with substantial top-up components and research support including equipment and travel funds.
Watch a brief video about my passion on bio-inspired robot navigation
Project Summary
This project aims to create fit-for-purpose positioning systems that continuously adapt to diverse and changing environments. The project expects to contribute to the knowledge across robotics, computer vision, and neuromorphic computing. Expected outcomes of this project include ground-breaking place recognition techniques that address two fundamental limitations in the state-of-the-art: continuous adaptation, critically important in safety-critical systems, and energy efficiency, critically important in resource-constrained systems. This should provide significant benefits, such as accelerated deployment of mobile robots, drones and augmented reality solutions in manufacturing, defence, healthcare, household, and space.
Motivation
Where are you? This is a fundamental question to which most of us usually know the answer. And so do the birds singing in our gardens, the aeroplanes flying over our cities, and the smartphones in our pockets. But why do we need to know where we are? Primarily because we need to navigate. To buy food, to go to work, to meet friends. And to navigate, we need to know where we are. And so do the robots and augmented reality devices of the future.
In addition, positional knowledge has widespread applicability beyond navigation. For example, it can help us and artificial intelligence make better decisions. Augmented reality devices can display location-aware content, interplanetary rovers can repeatedly sample the same locations, and we can better track and respond to natural disasters.
The importance of positioning or localisation systems has long been known in robotics. When robots localise visually, the problem is often called visual place recognition (VPR). Crucially, VPR enables navigation and decision-making without risky dependence on satellites, which are unavailable indoors, unreliable in built-up areas, and over which Australia has no control.
Aims
This DECRA fellowship will address three shortcomings of current VPR systems that have largely been overlooked but will be crucial for VPR’s widespread adaptation in edge devices:
- Current VPR systems are “one size fits all”. They do not adapt well to changing environments, like transitioning from rural to urban areas, which results in brittle solutions in environments that have not been encountered before.
- VPR algorithms do not exploit complementary information from non-positioning tasks that operate in parallel, like object detection and semantic segmentation, which in turn do not make use of positioning priors.
- There is a heavy focus on conventional computing, raising questions about long-term sustainability, data privacy in emerging applications, and ever-increasing power consumption. Neuromorphic computing provides a viable alternative for low-power edge devices with superior latencies, adaptability, and data efficiency.
Research Programs
Research Program 1: Heterogeneity, Adaptability, and Federation
It has long been known that different place recognition algorithms excel in different environments, and designing a system that works well in all environments is challenging. This research program will explicitly train networks to excel in specific challenges that are complementary to those challenges learnt by the other networks. Additionally, our design will allow for dynamically changing budget and accuracy requirements. Advances will include: 1) Heterogeneous models that are trained to perform well in specific types of environments and adapt to different challenges. 2) Predicting and forecasting the best-performing subset of models over time, such that not all models need to run at each time step. 3) Solving issues around data privacy using a distributed learning strategy.
Research Program 2: It’s Not All About Positioning
Current positioning techniques typically operate in a silo, that is, in parallel to, but independent from, other algorithms that solve related tasks such as object recognition and semantic segmentation. This research program aims to incorporate prior information from other systems to simplify positioning, enabling deployment on low-cost edge devices. Similarly, we will integrate place priors into non-positioning tasks to maximise performance. The research program will develop methods that jointly learn positioning & non-positioning tasks in a multi-task learning setting, which is advantageous as some concepts are easy to learn for one task while challenging to learn for another.
Research Program 3: Neuromorphic Place Recognition
Research Programs 1 and 2 introduce high-performing and adaptive place recognition algorithms deployed on conventional hardware. The third Research Program is complementary to those and will leverage advances in neuromorphic sensing and processing hardware to reduce latencies, increase data efficiency, and express uncertainties that edge devices can use. We will devise methods that dynamically adjust the hardware parameters of event cameras to optimise task performance, develop spiking neural network ensembles, and introduce a fully neuromorphic pipeline for visual place recognition.
Media and Links
Opportunities
To register your interest for one of the PhD positions or related positions, please e-mail your CV to tobias.fischer AT qut.edu.au.
Funding / Grants
- ARC DECRA (2024 - 2027)