Sudeep Dasari

I'm a PhD student at the Robotics Institute in Carnegie Mellon's School of Computer Science. I aspire to build scalable robotic learning algorithms, which can parse the visual world and enable autonomous agents to perform complex tasks in diverse environments. I am advised by Professor Abhinav Gupta.

In a prior life, I was an undergraduate student at UC Berkeley - where I work with Professor Sergey Levine on deep reinforcement learning/machine learning research. I also worked at Los Alamos National Laboratory with Dr. David Mascareñas on cyber-phsyical systems research.

Email  /  CV  /  GitHub  /  Google Scholar


I'm interested in computer vision, machine learning, robotics, statistics, and generative modelling. I aspire to push the boundaries of what robots can do with visual data given minimal supervision, and create new ways for them to interact more naturally with humans.

RoboNet: Large-Scale Multi-Robot Learning
Sudeep Dasari, Frederik Ebert, Stephen Tian, Suraj Nair, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Sergey Levine, Chelsea Finn
3rd Conference on Robotic Learning, 2019
project page, code, Press Coverage

We demonstrate that a large and diverse data-set of robotic experience (video + robot telemetry) can enable robot learning algorithms to learn new skills in new environments, faster than training from scratch.

Visual Foresight: Model-Based Deep Reinforcement Learning for Vision-Based Robotic Control
Frederik Ebert*, Chelsea Finn*, Sudeep Dasari, Annie Xie, Alex X. Lee, Sergey Levine
arXiv Preprint, 2018
project page, code, blog post

We demonstrate that video prediction models - generative models trained on autonomously collected data of robots interacting with their environment - can be used for vision based robotic control.

Robustness via Retrying
Frederik Ebert, Sudeep Dasari, Alex X. Lee, Sergey Levine, Chelsea Finn
2nd Conference on Robotic Learning, 2018
project page, code

While visual prediction planners can acheive a wide variety of manipulation tasks, they often fail by loosing track of the object. We propose a self-supervised registration algorithm which allows even imperfect planners to retry continuously until they succeed.

Domain Adaptive Meta-Learning
Tianhe Yu*, Chelsea Finn*, Annie Xie, Sudeep Dasari, Tianhao Zhang, Pieter Abbeel, Sergey Levine,
Robotic Science and Systems, 2018
project page, code

Domain Adaptive Meta-Learning allows for one-shot learning under domain shift. Using our policy, robots can to learn to manipulate unseen objects by referring to a single video demonstration of a human performing a task with said object.

Deep Robotic Learning using Visual Imagination & Meta-Learning
Chelsea Finn, Sudeep Dasari*, Annie Xie*, Frederik Ebert, Tianhe Yu, Pieter Abbeel, Sergey Levine
Demonstration at Neural Information Processing Seminar, 2017
project page, press coverage

A key, unsolved challenge for learning with real robotic systems is the ability to acquire vision-based behaviors from raw RGB images. We present two approaches to this goal that we demonstrate: first, learning task-agnostic visual models for planning and second, learning to quickly adapt to new objects and environments using meta-imitation learning.

Light Field Imaging of Three-Dimensional Structural Dynamics
Benjamin Chesebrough, Sudeep Dasari, Andre Green, Yongchao Yang, Charles R. Farrar , David Mascareñas
Society for Experimental Mechanics: IMAC, 2018

Cameras offer a way to perform low cost and non-invasive damage detection by analyzing a structure's vibrations. To acheive this goal, we use a light field imager to create a 3D point cloud of vibrating cantilever beams, and then use Principal Component Analysis and Blind Source Seperation to find the beam's mode shapes and frequencies.

A framework for the identification of full-field structural dynamics using sequences of images in the presence of non-ideal operating conditions
Sudeep Dasari, Charles Dorn, Yongchao Yang, Amy Larson , David Mascareñas
Journal of Intelligent Material Systems and Structures, 2017

Recent developments in analyzing structures with raw video streams has great potential for reducing the resources and time required for performing operational modal analysis at very high spatial resolution. We propose a key-point based registration technique to extend this method to settings with large rigid body motion (e.g drone footage).

Efficient Full-Field Vibration Measurements and Operational Modal Analysis Using Neuromorphic Event-Based Imaging
Charles Dorn, Sudeep Dasari, Yongchao Yang, Charles R. Farrar , Garrett Kenyon, Paul Welch, David Mascareñas
Journal of Engineering Mechanics, 2017

We demonstrate that asynchronous silicon retina imagers can capture structural motion on the microsecond scale in an extremely data-efficient manner.

Course Projects

Fun With Words: Generating Text with Context
Sudeep Dasari, Hankun (Jerry) Zhao, William Zhao, 2018

State of the art machine translation networks can be used for conditional text generation. We leverage the Yelp dataset to train models which generate sensible 1 - 5 star reviews of resteraunts given only the establishment's name.

Non-Linear Raytracing
Sudeep Dasari, Jennifer Wen Hsu, 2017

Most ray-tracing algorithms assume that light travels in a straight line from the camera into the scene. However, there are many cases where a light may travel a non-linear path (e.g mirage). We demonstrate non-linear raytracing can correctly render scenes with mirages and fata-morgana, as well as trace light passing through Gradient Optic Index Lenses.

Template courtesy of here