Project 11

Project location:

Primary location: Pullenvale (QLD); other potential locations include Black Mountain (ACT) and Clayton (VIC).

Desirable skills:

1. Strong programming skills with C++ in Linux and experience with ROS (RP1), strong programming skills with Python and  PyTorch (RP2 and RP3).

2. Strong mathematical skills for all projects. Prior experience and background in (numerical) linear algebra, statistical inference, and/or mathematical optimisation is highly desirable.

3. Familiarity with SLAM and GNSS signal processing (RP1 and RP2). Understanding of SLAM and self-supervised learning, especially transformers, (RP2 and RP3).

4. Strong oral and written communication skills, good problem-solving skills and ability to work on applied research problems. 

Supervisory project team:

Milad Ramezani, Reza Arablouei, Kasra Khosoussi, Zeeshan Hayder and Moshiur Farazi

Contact person:

Project description:

Outline:
Autonomous experimentation, in a collaborative fashion between robots and a team of scientists, requires data analysis in a global coordinate system. To this end, robust and versatile localisation and mapping in a geographic coordinate system are key components. For example, an autonomous agent (e.g., vehicle or robot) requires accurate and seamless Geo-localisation when transitioning across various environments (e.g., from an open sky area where GNSS is the only accessible data to unstructured regions with limited sky view such as under the canopy or to structured environments such as urban areas including tunnels) and transferring information. Global navigation satellite system (GNSS) receivers offer valuable geotagged positional measurements outdoors and significantly mitigate long-term drift due to accumulated errors in simultaneous localisation and mapping (SLAM) in the absence of loop closures. However, GNSS measurements are susceptible to corruption by various error sources such as multipath and atmospheric interference. This makes fusing GNSS data with other sensor information non-trivial, as it requires careful consideration of the uncertainties and challenges associated with GNSS measurements. Many scientific experiments and tasks require operation in challenging GNSS-denied environments such as mines and caves. These environments are often prone to sensing degeneracy due to their self-symmetry and featureless nature. Existing SLAM systems lack effective mechanisms for detecting and handling the resulting ill-conditioned estimation problems and thus may fail in such challenging situations.

This proposal aims to overcome the limitations of the existing SLAM solutions by developing a robust and versatile SLAM approach in a geographic coordinate system. The solution will leverage GNSS signal processing, self-supervised learning, and the well-established theory of factor graph. The research will be carried out through three synergistic PhD projects, each pursued by one student.

– Two PhD projects will address the challenges of incorporating raw GNSS data into SLAM in environments where GNSS signals are available or degraded. The research in these projects will be conducted from two perspectives
1) to provide a seamless localisation and mapping solution with centimetre-level accuracy leveraging GNSS signal processing and factor-graph optimisation; 
2) to model the characteristic behaviour of GNSS errors, such as multipath, in various environments utilising self-supervised deep learning techniques and employ the learned models for calibrating GNSS signals. 

– The third PhD project
3) aims to investigate the issue of sensing degeneracy in SLAM that frequently arises in challenging GNSS-denied environments (e.g., tunnels), develop novel methods for detecting degenerate situations, and design and implement effective mitigation strategies for robust localisation and mapping performance under sensing degeneracy.

A collaborative and knowledge-sharing environment will be fostered among the students, encouraging them to work together and build upon each other’s research. This will stimulate creativity, innovation, and cross-disciplinary learning, enhancing the overall progress and success of the research endeavour.

Activities the students will undertake:
The PhD students will work on three distinct Research Packages (RPs), each contributing to the overall research objectives of the proposal. Their individual efforts and expertise will collectively advance the development of a robust and versatile SLAM solution. In particular, the activities of each PhD project are summarised as: 

RP1: Incorporating raw GNSS measurements, i.e., pseudo ranges and carrier phases, into SLAM factor graph. The student will develop a customised factor for GNSS raw measurements to add into a factor-graph SLAM pipeline in conjunction with other factors (i.e., lidar and IMU). The student will conduct a comprehensive literature review of GNSS principles and factor-graph SLAM. They will collect data, devise innovative solutions, and perform experimentation to validate and refine the proposed models and techniques. The research outcomes will be published in top-tier robotic and geoscience venues.

RP2: Data-driven calibration of raw GNSS data. The student will utilise self-supervised deep-learning techniques to build foundational encoder-decoder models that can characterise and predict the errors of GNSS signals in various environments. The student will cover the literature review related to deep-learning techniques with the focus on GNSS signal processing. They will collect a large GNSS dataset in different areas with various GNSS error characteristics to train/tailor the network. The research outcomes will be published in top-tier robotic and CV venues.

RP3: Developing novel estimation-theoretic (e.g., observability analysis) and learning-based methods for detecting sensing degeneracy in SLAM that can lead to ill-conditioned/posed estimation and optimisation problems. The student will develop novel mitigation strategies based on adaptive sensor fusion (IMU, lidar, and vision) and/or non-Gaussian inference (to capture and keep track of ambiguity in measurements and the belief) for robust localisation and mapping under sensing degeneracy. They Implement the proposed techniques and algorithms on the state-of-the-art SLAM systems and collect data and conduct experiments in simulated and real environments to validate the proposed techniques and evaluate their performance. Research outcomes will be published in top-tier conferences and journals.

Expected outcomes:
The nominated candidates work effectively together, and in collaboration with the university supervisors the expected outcomes are: 

– Development of a lidar SLAM software package by introducing novel GNSS constraints in optimisation to achieve centimeter-level accuracy for robot pose estimation in real-time. The trained GNSS model is integrated with the GNSS-Lidar SLAM solution to calibrate GNSS measurements in real-time in GNSS-degraded environments. The student will gain solid knowledge of GNSS and SOTA deep learning algorithms and develop skills in programming, academic writing and collaboration with researchers and engineers.

– Development of a novel ML approach to calibrate GNSS raw measurements in different environments, such as urban canyons and under the canopy. In inference, the trained model allows to mitigate errors (mainly multipath) from GNSS measurements. The student will gain solid knowledge of GNSS and SLAM and develop skills in programming, academic writing and collaboration with researchers and engineers.

– Design and implementation of novel mathematical and data-driven methods for detecting and mitigating sensing degeneracy using multiple sensing modalities (e.g., lidar and vision) and their integration with state-of-the-art SLAM systems.  This will enable robust and resilient localisation and mapping in some of the most challenging GNSS-denied environments such as underground tunnels. The student will have the opportunity to learn about various subjects such as robot perception, deep learning, mathematical optimisation, estimation theory and observability analysis. They will also develop their programming and communication skills, and experience working in teams with other students, CSIRO researchers, and their academic supervisors on cutting-edge problems in robotics and AI.

The outcomes of this project can potentially be used for monitoring assets such as forests or coral reefs and the aquatic species in the oceans aligned with CSIRO missions such as drought resilience and ending plastic waste.