Nishad Gothoskar

I am a first-year PhD Student at MIT where I work in the Probabilistic Computing Group and am advised by Vikash Mansinghka and Josh Tenenbaum. I am interested in building intelligence for robots. By combining insights from the cognitive and neuro sciences with probabilistic generative models and Bayesian inference, I seek to design algorithms that can match humans' ability to learn and generalize.

For the years before I joined MIT, I was a research engineer at Vicarious AI where I was advised by Dileep George and Miguel Lázaro-Gredilla. My research was on sequence learning models, approximate inference, and generative models applied to robotics. I have also worked for short periods at Uber ATG, Google, and MIT. I received my undergraduate degree from Carnegie Mellon, where I studied math and computer science.

Research
From proprioception to long-horizon planning in novel environments: A hierarchical RL model
N. Gothoskar, M. Lázaro-Gredilla, D. George
arXiv

In this work, we introduce a simple, three-level hierarchical architecture that reflects these distinctions. The low-level controller operates on the continuous proprioceptive inputs, using model-free learning to acquire useful behaviors. These in turn induce a set of mid-level dynamics, which are learned by the mid-level controller and used for model-predictive control, to select a behavior to activate at each timestep. The high-level controller leverages a discrete, graph representation for goal selection and path planning to specify targets for the mid-level controller.

Query Training: Learning and inference for directed and undirected graphical models
M. Lazaro-Gredilla, W. Lehrach, N. Gothoskar, A. Dedieu, G. Zhou
arXiv

We introduce Query Training (QT), a systematic method to turn any PGM structure (directed or not, with or without hidden variables) into a trainable inference network. This single network can approximate any inference query. We demonstrate experimentally that QT can be used to learn a challenging 8-connected grid Markov random field with hidden variables and that it consistently outperforms the state-of-the-art AdVIL when tested on three undirected models across multiple datasets.

Uncalibrated Visual Servoing
N. Gothoskar, M. Lazaro-Gredilla, A. Agarwal, Y. Bekiroglu, D. George
arXiv

Developed a method for learning to control a robot using visual feedback with no calibration or prior information about the setup.

Learning cognitive maps from vicarious trial and error
R. Rikhye, N. Gothoskar, S. Guntupalli, A. Dedieu, M. Lazaro-Gredilla, D. George
bioRxiv

Cognitive maps enable us to learn the layout of environments, encode and retrieve episodic memories, and navigate vicariously for mental evaluation of options.

Learning higher-order sequential structure with cloned HMMs
A. Dedieu, N. Gothoskar, S. Swingle, W. Lehrach, M. Lazaro-Gredilla, D. George
arXiv

CHMMs are a constrained version of HMMs with a simple sparsity structure that enforces many hidden states to map deterministically to the same emission state. We demonstrate the effectiveness of this model in various data domains.

Predicting interactions for autonomous vehicles
Uber Advanced Technologies Group

Worked on road user trajectory prediction to enable planning. I built system for modeling interactions between road users.

Does this App Really Need My Location?
S. Chitkara, N. Gothoskar, S. Harish, J. Hong, Y. Agarwal
Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies (ACM IMWUT)
[Website] [Press] [Interview]

Mobile applications prey on user's sensitive data. We developed an application that monitors the flow of private information and protects their data from being leaked.


Template from here.