Stein Particle Filter for Nonlinear, Non-Gaussian State Estimation | Robotics for Climate Change Workshop @ ICRA 2022

50 views
Download
  • Share

About the Video:
The sequential estimation of posterior distributions in state-space models from noisy observations is a critical capability in robotics, particularly in tasks such as localization, object pose estimation and terrain modelling. While sequential Monte Carlo techniques such as particle filters provide a solution for nonlinear dynamics and observation models, they do not scale well with the dimensionality of the problem. This talk introduces a new filtering technique that can leverage differentiability in both observation and dynamics models to produce nonparametric, non-Gaussian posteriors. As with particle filters, the posterior is represented by a set of particles but follow a different update rule based on Stein variational gradient descent that minimizes the Stein discrepancy. This talk shows that Stein Particle Filters scale better with the dimensionality of the data and can be implemented efficiently on GPUs. Other methods based on Stein variational gradient descent for control and motion planning are also discussed.

About the Speaker:
Fabio Ramos is a Principal Research Scientist at NVIDIA, and Professor in machine learning and robotics at the School of Computer Science, University of Sydney. His research interests are in probabilistic inference and robotics.

Recorded at the Robotics for Climate Change Workshop during the 2022 IEEE International Conference on Robotics and Automation (ICRA).

Produced in partnership with the IEEE Robotics and Automation Society (https://www.ieee-ras.org/).

Recording funded in part by a grant from the United Engineering Foundation (https://www.uefoundation.org/).
 

About the Video:
The sequential estimation of posterior distributions in state-space models from noisy observations is a critical capability in robotics, particularly in tasks such as localization, object pose estimation and...

Advertisment

Advertisment