Riley Edmunds

At UC Berkeley, I researched deep reinforcement learning for robotics in Alice Agogino's lab and adversarial machine learning in Dawn Song's lab at BAIR. My robotics work included learning policies for tensegrity robots developed with NASA Ames.

I contributed to research published at ICRA, IROS, and DLSW, and co-wrote a chapter on adversarial machine learning in Artificial Intelligence Safety and Security. I also built and led the research arm of Machine Learning at Berkeley.

At Bridgewater Associates, I was part of the firm's early AI lab, researching machine learning methods for macro investing. I was also a Jane Street Fellow at Depth First Learning, where I co-wrote and taught a course on the theory of deep neural networks.

Most recently, I co-founded Stride Labs, where we developed cross-chain infrastructure and built and operated our own blockchain. We raised $15M and grew to over 100K users.

Email  /  GitHub  /  LinkedIn  /  Scholar

Developed and tested locomotion strategies for tensegrity robots in a joint Berkeley–NASA research effort. Demonstrated reliable climbing on inclines up to 24 degrees using new multi-cable actuation strategies.

Demonstrated that tensegrity robots can learn locomotion policies even with severely limited sensory inputs using Mirror Descent Guided Policy Search (MDGPS). Evaluated in simulation, with preliminary tests on physical hardware.

Co-wrote and taught a class on deep neural network signal propagation via random matrix theory as a Jane Street Fellow. Depth First Learning is a collaboration between Google Brain, DeepMind, FAIR, and NYU.

Explored whether autoencoder reconstruction errors could predict transferability between reinforcement learning tasks. Evaluated the approach on MuJoCo Ant locomotion tasks and experimented with integrating it into Model-Agnostic Meta-Learning (MAML). Also built a MuJoCo simulation environment for tensegrity robot locomotion.

Studied the transferability and generalization of deep neural network policies. Tested this by seeing whether adversarial attacks carry from a shared Model-Agnostic Meta-Learning (MAML) model to models fine-tuned for different tasks.

A chapter surveying adversarial examples in deep networks: how gradient-based attacks craft input perturbations that flip a model's prediction, how those attacks transfer across architectures, and what that implies for the robustness of deployed systems.

One of my first research reports at Bridgewater. Analysis of the unprecedented speed and magnitude of policy response to the COVID-19 crisis, comparing the timeline of asset reflation across the Great Depression, Global Financial Crisis, and COVID-19 pandemic.

A hierarchical model combining a semi-supervised autoencoder with a supervised classifier for fraud detection in data with imbalanced classes and concept drift. This work resulted in U.S. Patent 11829866, granted in 2023.