Abstract:

Autonomous robots shall soon bring AI to the physical world, be it by automating agriculture, cleaning or servicing homes and hospitals, or studying life in the depths of the oceans. As data science is a key enabler for translating the robot sensor inputs into actionable information for decision-making, efficient and reliable data analysis is needed to design such efficient, reliable, and adaptable robots.  Pose estimation is a critical data analytics challenge in robotics: to determine a robot’s precise location and orientation in dynamic and unstructured environments using sensors. Accurate pose estimation is critical for autonomous navigation, enabling robots to make informed decisions and mitigate risks. However, existing methods, particularly deep learning-based, suffer from overconfidence in their predictions, leading to unreliable decision-making under uncertainty.
I will adapt the novel Stein mixture inference sequential modeling to address significantly improve pose estimation. Existing Bayesian inference techniques for pose estimation are either inaccurate or inefficient.  Stein mixture inference combines the flexible representation of particle variational methods with the broad coverage of mixture models, offering a computationally efficient, accurate, and uncertainty-aware representation of robot pose. I will focus on two key objectives:
1. Exploring the applicability of Stein mixture inference to pose estimation, including adapting Stein inference to pose estimation and experimenting with lidar, radar, and stereo-camera modalities, benchmarking against deep learning-based methods. I hypothesize that Stein inference will achieve comparable accuracy while mitigating the overconfidence pitfalls of deep learning models, improving the reliability of autonomous navigation.
2. Ensuring real-time feasibility by performing runtime and memory analysis to validate computational efficiency. I hypothesize that introducing nonlinear optimization will achieve real-time time processing performance at an improved convergence rate.
Achieving this breakthrough requires a close collaboration of an experienced data-science researcher with the robotics environment, combining the excellence in Bayesian inference with access to robotics expertise and infrastructure. Together we can bring the future of autonomous agriculture, deap ocean and climate science research closer.