Nitish Dashora

I am a post-bacc at UC Berkeley where I am a researcher at Berkeley Artificial Intelligence Research under Prof. Sergey Levine. I used to conduct research under Prof. Bruno Olshausen from the Redwood Center for Theoretical Neuroscience. Lastly, I was a member of Machine Learning at Berkeley where I led research projects, course offerings, and mentor incoming students. I recieved my B.S. in Electrical Engineering and Computer Science (EECS) from UC Berkeley.

Summer 2023: I was a research intern at the Center for Human-Compatible AI working under Prof. Stuart Russell to study human-inspired exploration and skill discovery paradigms.
Summer 2022: I was a software development engineer intern at Amazon working with Amazon Web Services to create optimized Spark/Hadoop algorithms that use multiple clusters from load balancing across sharded data.
Summer 2021: I was a research fellow at NASA JPL where I worked with team CoStar, a joint MIT-Caltech-NASA organization for robotics research. I specifically worked with skill development for robotic planning through deep learning techniques.
Fall 2021: I was a Schmidt Futures AI Fellow for Educational AI Technology and entrepreneurship.

Email  /  CV  /  Bio  /  Twitter  /  Github  /  LinkedIn

I am most interested in the theory of intelligence and how to create general learning systems that encompass this definition. The age-old question of AGI has so strongly captivated me that I am certain I would like to dedicate my life’s career to exploring it. I think that physical embodiment will play a significant role in the creation of AGI, thus I find that robotics and robot learning is an excellent testbed for AI research. Currently, I work in computer vision and reinforcement learning to create rich perceptual learning systems and decision-making agents. My current research under Prof. Sergey Levine at BAIR involves zero-shot reward shaping for online RL and subgoal reaching. I also have interests in representation learning which I pursued with Prof. Bruno Olshausen by exploring object-part relationships and deformation through Restricted Boltzmann Machines.

We present a compositional skill discovery feedback loop for fast online exploration and planning.

We present a multi-robot visual navigation foundation model for effective fine-tuning and novel task adaptation.

DAgger-like imitative learning with model predictive path integral control works as an end-to-end navigation system.

Combining geometric costmaps and learned models improves long-range navigation in offroad environments.

source