End-To-End Learnable Histogram Filters

Betreuer: Heiko Donat

A classical approach to solve problems in robotics is to develop problem-specific algorithms which include prior knowledge about the robot and the environment to find robust and data efficient solutions.
Another approach which gains popularity in robotics is the usage of machine learning that permits adaption to the task without the need of a human expert. Furthermore, it is possible to solve problems which are currently infeasible to solve by classical engineering.
Recently, there is an increased interest to combine both approaches to balance strengths and weakness.

One approach shown in the paper End-To-End Learnable Histogram Filters by Rico Jonschkowski and Oliver Brock presents a differentiable implementation of a histogram filter (which is used for robot localization and represents the robot state in form of a histogram).
By substituting parts of the histogram filter by deep neural networks, it is possible to optimize the performance of the filter by supervised or unsupervised learning.
In this seminar talk should present the paper and its related work.

Possible Sources:

  1. Robot Localization II: The Histogram Filter
  2. Towards Combining Robotic Algorithms and Machine Learning:
    End-To-End Learnable Histogram Filters
  3. Backprop KF: Learning Discriminative Deterministic State Estimators