I’m Samuel Pfrommer, a PhD student at Berkeley EECS working with Somayeh Sojoudi. My research interests span robust machine learning, safe reinforcement learning, and geometric deep learning.




Google Scholar

Curriculum Vitae


GNN Discriminability. Analyzed the impact of nonlinearities on the discriminability of single-layer graph neural networks. Proved theoretical results showing that nonlinearities make GNNs more strictly more discriminative than regular graph filters in practical settings.

ContactNets. Reparameterized the contact dynamics learning problem to handle nonsmooth impact and stiction for a real-world block toss.

Slipping SLIP. Used numerical trajectory optimization to look at Spring Loaded Inverted Pendulum traversal strategies for slippery terrain.

Project samples

Airfoil pressure field learning. Used CNN approaches to learn approximations of slow CFD pressure field computations.
MATLAB terrain elevation plugin. Wrote an easy-to-use MATLAB plugin that fetches, combines, and renders USGS terrain elevation data.

Airfoil configuration optimization. Optimized the placement of several airfoil elements on a racecar rear wing using genetic algorithms.

Legged helicopter landing gear. Designed and built a legged landing gear for my RC helicopter, with spring-loaded, force-sensitive legs and retraction reflexes for landing on uneven terrain.

Homemade birthday gifts

My friend and roommate is really into music and photonics. So I collaborated with a mutual friend to build him a music visualizer, largely based on this project. The oak base is entirely cut and mitered by hand, and we soldered the same circuit used in the instructable to make it sound reactive. The final project has five LED-lit vacuum tubes, a microphone, and switches to toggle power and sound reactivity.
The friend who helped me build the vacuum tube project had the next birthday. He loves whiskey and beagles, so we decided make him a beagle ice cube mold. We got a beagle keychain to make the negative using a silicone mold making kit. Even though our viscosity was way too high the molds still ended up being mostly bubble free!
My girlfriend is a fan of plushies and Rubik’s cubes, so I combined the two. I originally tried gluing the tiles but the epoxy discolored the fabric. Sewing wasn’t nearly as bad as anticipated though.


  1. Hi Sam,

    My Name is Matt Kelly – you link to my website on your trajectory optimization page. These projects are all really impressive – nice work. If you’re interested, send me an email and we can set up some time to chat on the phone or skype. I’m curious to hear what you’re working on next! Best wishes,



  2. Hi Sam, I’m Sam too.
    Regarding your work “Airfoil Configuration Optimization with Genetic Algorithms”, could you release the coordinate geometries for the wing/aerofoils you trialed?

    I’d love to read your paper on this project and possibly use it to develop my own iterative development of multi-element wings

    Samuel, CEng MEng


    1. Hi Sam,

      I used the s1223 airfoil since that’s what we were using on the car.

      If you are curious in trying around with different airfoils the UIUC database is a good start.



Leave a Reply

Your email address will not be published. Required fields are marked *