About Me

Craig Knuth
Master’s in Robotics
University of Michigan

I am a Master’s candidate in Robotics at the University of Michigan. I work in the ARM Lab under Dmitry Berenson and in Necmiye Ozay’s research group.


  • Mobile robotics
  • Safety under scarce information
  • Stochastic modeling and control
  • Safe learning-based control


  • Master’s in Robotics,University of Michigan
  • Bachelor’s in Computer Science, UW-Madison
  • Bachelor’s in Mathematics, UW Madison
  • Certificate in Physics, UW-Madison

Research Projects

Cell decomposition of the space based on the agent’s vision model.

Learning Obstacles From Trajectories of 2D Vision-Based Navigators

If we only have access to the trajectories of autonomous agents, can we infer a likely structure of obstacles in the environment? We also assume that agents have no prior knowledge of the environment and only act based on what they can see and what they have seen. In this work we “imagine” obstacles that we know would have been seen from at least one point in the demonstration and check its consistency with the trajectories.

“Inferring Obstacles and Path Validity from Visibility-Constrained Demonstrations.” C. Knuth, G. Chou, N. Ozay, D. Berenson. Submitted to RA-L September 2019.

Evaluation of Control Methods Paired with Learned Methods

The world is now filled with models based on data which forgoes many guarantees that traditional control methods provide. These models often achieve success when paired with an appropriate control strategy, such as model predictive control. We explore how much success we should attribute to the learned model versus the particular control method, as well as how we improve the overall resulting systems.

Applying Deep Unsupervised Learning to Robotic Manipulation

Deep unsupervised learning seeks to fit a probability distribution to a dataset in order to evaluate the probability of any particular novel data point or sample new data points. In robotic manipulation of deformable objects, we cannot plan using existing models as they are too computationally intensive. By applying unsupervised learning techniques, we seek to capture the model in a rich enough way that we can utilize existing methods of planning in uncertain environments.

Model of pest generation dynamics

Stochastic Simulation and Analysis of Pest Resistance Evolution

Many models of pest resistance evolution are deterministic due to the difficulty in analyzing analogous stochastic formulations. In agriculture, these models give confidence to farmers that applying pesticide while maintaining a refuge field will slow evolution of pesticide resistance. However, with more accurate stochastic models it may be that resistance evolves more quickly. This work seeks to establish and analyze an appropriate stochastic model to explore this phenomena.

Contact Me