Somnath Kumar

Somnath Kumar

Research Fellow, Microsoft Research India

Indian Institute of Technology (BHU), Varanasi

Biography

I am a Research Fellow at Microsoft Research working on alignment and optimization of language models using different Reinforcement learning algorithms. My interests encompasses but not limited to Representation learning and Embodied AI. My research work has been in developing systems and agents which benefit or rely on these algorithms.

I have joined John Dolan’s Argo AI Lab at Robotics Institute, Carnegie Mellon University working on Safe Autonomous Navigation.

Actively working under Dr. Shishir Kolathaya, IISc, bangalore at Stochlab on Different Reinforcement Learning Algorithm for Quadrupeds.

Previsously I worked under Dr. Pratik Chattopadhyay developing Generative Adverserial Networks to generate occluded pose in the whole Gait Cycle of a human subject.

Besides Floating point values I like strings on my guitar to play a few tunes.

Download my resumé.

Interests
  • Reinforcement Learning
  • Representation learning
  • Computer Vision
  • Optimal Control
Education

  • BTech in Electrical Engineering, 2019-2023

    Indian Institute of Technology (BHU), Varanasi

News

Recent Publications

LatentCBF: Control Barrier Definition on Latent Space
Safe execution of policies is crucial for an autonomous system. A popular choice for safety guarantees is CBF(Control Barrier …
Somnath Kumar, Qin Lin, John M. Dolan
PDF
LatentCBF: Control Barrier Definition on Latent Space
Realizing Linear Controllers for Quadruped Robots on Planetary Terrains
Until now, planetary exploration has been accomplished with wheeled vehicles, which makes movement in highly complex, sandy, and …
Aditya Shirwatkar, Somnath Kumar, Bharadwaj Amrutur, Shalabh Bhatnagar, Ashitava Ghosal, Shamrao, Vinod Kumar, Shishir Kolathaya
PDF Video
Realizing Linear Controllers for Quadruped Robots on Planetary Terrains
BGaitR-Net: Occluded Gait Sequence reconstruction with temporally constrained model for gait recognition
Recent advancements in computational resources and Deep Learning methodologies has significantly benefited development of intelligent …
Somnath Kumar, Pratik Chattopadhyay, Lipo Wang
PDF
BGaitR-Net: Occluded Gait Sequence reconstruction with temporally constrained model for gait recognition

Experience

 
 
 
 
 
Data Science Intern
Bosch AI Sheild Department
May 2022 – Jul 2022
Worked on Vulnerability Analysis for Video Classification Model against Black Box Extraction.
 
 
 
 
 
Research Assistant
Carnegie Mellon University
Mar 2022 – Present
Working on Safe Reinforcement Learning for autonomous driving.
 
 
 
 
 
Research Intern
Stochlab, IISc.
Apr 2021 – Jun 2022
My role included developement of ROS package for the quadruped named Stochlite, for deploying optimal control algorithms on it. On the Later Stage worked towards model based learning methods for robust walking on many challenging terrains.
 
 
 
 
 
Research Intern
Pattern Recognition Lab, IIT (BHU), Varanasi.
Dec 2020 – Mar 2021
Worked on developing gait occlusion reconstruction algorithms using Conditional Variational Auto encoder and LSTMs for reconstructing the occluded gait cycle, Also worked on implementing other algorithms for comparative study.

Projects

*
All MultiAgent Reinforcement Learning ROS Robotics Other
ZeroxQA

ZeroxQA

Few shot learning for QA tasks on outdomain queries

Slides Code and Details
Black Box Model Extraction

Black Box Model Extraction

Extraction of Model using Black box queries, from limited or no original data.

Code Slides
Graph Neural Network for Communication in MARL

Graph Neural Network for Communication in MARL

Enabling communication in MultiAgent Reinforcement Learning using Graph.

Code and Details
MultiAgent Full coverage

MultiAgent Full coverage

efficient coverage using multiple agents in unknown terrain

Code and Details Abstract
NanoDrone RL Aided MPC

NanoDrone RL Aided MPC

Reinforcement Learning aided MPC for constrained systems

Hardware
Mini Swarm Bot

Mini Swarm Bot

Swarm of robots for warehouse management

Code and Details
Optimal control and trajectory optimization for Quadruped

Optimal control and trajectory optimization for Quadruped

Different control and optimization algorithms on Stochlite, IISc.

Code (Private Repository)
Hierarchical Reinforcement Learning for Swarm of Modular Bot

Hierarchical Reinforcement Learning for Swarm of Modular Bot

Enabling communication in MultiAgent Reinforcement Learning using Graph.

Gait-Occlusion-Reconstrution

Gait-Occlusion-Reconstrution

Reconstruction Gait sequence for gait recognition

Code and Details
PauciBot

PauciBot

Working around the 2 wheels.

Code and Details
KiloBot MultiAgent Reinforcement Learning

KiloBot MultiAgent Reinforcement Learning

Deep MultiAgent Reinforcement Learning on “KiloBot” by Havard University

Code and Details
Hand Imitation

Hand Imitation

Reinforcement Learning for an Imitating Robotic Arm.

PDF Code and Details Abstract
Pong Reinforcement Learning

Pong Reinforcement Learning

Implementation of Policy Gradient on self made Pong game.

Code and Details
Light Gun

Light Gun

Revival of our beloved video game with Computer Vision.

Code and Details

Popular Topics

3D Modelling Academic Computer Vision Control Theory Deep Learning Embedded System Embedded Systems Funds Graph Neural Network Linux Model Extraction MultiAgent MultiAgent Systems Navigation Stack OpenAI gym Optimal Control PyGame Reinforcement Learning Robotics ROS

Contact