My Profile Photo

Sebastian Blaes


Ph.D. candidate in the Autonomous Learning Group at the Max Planck Institute for Intelligent Systems, Tuebingen.


Cycling, running, trekking, mountaineering enthusiast.


RAZER @ CoRL'21


RAZER @ CoRL'21

We are very happy to present our work Risk-Averse Zero-Order Trajectory Optimization at CoRL’21 in London. Our first in-person conference in a long time.

In this work we design a simple method to separate epistemic and aleatoric uncertainties in a learned ensemble of probabilistic forward dynamics models. We then incorporate these uncertainties explicitly in the cost function of our iCEM trajectory planner to control robotic agents in a MPC fashion.

Our key contributions are:

  • A new architecture for separating uncertainties in probabilistic ensembles.
  • Efficient use of epistemic and aleatoric uncertainties.
  • An explicit use of uncertainties in the cost, allowing for trading off uncertainty-aware and task-aware planning.
  • A simple but practical approach to probabilistic safty constraints.

Check out our poster below and our paper.

Resources

Poster

Poster

Poster


Videos

Solo8-LeanOverObject


Solo8-LeanOverObject