Mr Mingda Xu

I am a Research Fellow working in Prof. Stephen Gould's group at the Australian National University on a collaborative project with research scientists at industry partner Seeing Machines Ltd. I love working across the full robot "stack" and learning about all aspects of autonomous systems. Interesting machine learning and computer vision problems we have worked on in this project:

• Unsupervised learning for long-form videos (CVPR 2024 best paper award nominee)
• Feature shaping methods for out-of-distribution detection (ICLR 2024).
• Unsupervised anomaly detection using deep density estimation.
• Diffusion models for probabilistic pose estimation.

I also have projects running around model-based control, specifically discrete-time optimal control as well as differentiable optimisation for robotics. Examples of projects in this area are

• Differentiating through optimal control problems to learn models from sparse demonstrations (NeurIPS 2023).
• Differentiable Dynamic Time Warping (DTW) for sequence-based visual place recognition (ICLR 2023).
• Differential Dynamic Programming (DDP) with constraints.

My PhD looked at efficient visual localisation algorithms for mobile robots robust to appearance and viewpoint changes (long-term autonomy). We combined state-estimation algorithms based on recursive Bayesian filtering and smoothing and nonlinear optimisation with modern deep learning methods in visual place recognition.