Yide(Fred) Shentu

fredshentu at berkeley dot edu

I am a PhD student in the field of machine learning and robotics at UC Berkeley, advised by Professor Pieter Abbeel. Previously, I graduated from UC Berkeley majoring in Computer Science math and Physics. During my undergraduate study, I worked as an undergraduate research assistant in Berkeley Artificial Intelligence Research (BAIR) Lab. Post-graduation, I had the opportunity to be a founding engineer at Covariant AI. Witnessing the company's growth from an 8-person team to an almost 200-person enterprise spanning multiple continents was truly remarkable. Now, I have returned to academia to delve deeper into research.

Email  /  CV  /  Github

My research interest lies at the intersection of robotics, computer vision, and machine learning. I'm interested in developing algorithms which can make use of rich sensory data such as vision to not only make robot be able to complete complicated tasks but also improve themselves continuously by interacting with the environment.

From LLMs to Actions: Latent Codes as Bridges in Hierarchical Robot Control

we present Latent Codes as Bridges, or LCB, a new policy architecture for control that combines the benefits of modular hierarchical architectures with end-to-end learning

GELLO: A General, Low-Cost, and Intuitive Teleoperation Framework for Robot Manipulators

We propose GELLO, a general framework for building low-cost and intuitive teleoperation systems for robotic manipulation.

Autoregressive Uncertainty Modeling for 3D Bounding Box Prediction

We propose methods for leveraging the autoregressive model to make high confidence predictions and meaningful uncertainty measures for 3d bounding box prediction.

Learning Instance Segmentation by Experimentation

We present a robotic system that learns to segment its visual observations into individual objects by experimenting with its environment in a completely self-supervised manner.

Zero-Shot Visual Imitation

We present a novel skill policy architecture and dynamics consistency loss which extend visual imitation to more complex environments while improving robustness.

Probabilistically Safe Policy Transfer

We formalize the idea of safe learning in a probabilistic sense by defining an optimization problem: we desire to maximize the expected return while keeping the expected damage below a given safety limit.