Soumojit Bhattacharya

I am currently a Master of Science in Robotics (MSR) student at Carnegie Mellon University, advised by Professor Shubham Tulsiani and Professor Oliver Kroemer. My primary research focuses on robotic manipulation, developing manipulation pipelines that generalize across diverse objects and execute complex tasks by decomposing them into simpler motions.

Currently, I am working on diffusion policies for robotic manipulation, aiming to advance adaptable and robust manipulation strategies capable of handling variability in objects and tasks. My research strives to enable intelligent robotic systems that perform a wide range of challenging manipulation tasks efficiently and flexibly.

I graduated with a B.Tech. in Electronics and Electrical Communication Engineering from Indian Institute of Technology Kharagpur, where I worked with Professor Aritra Hazra. I was also a member of the Autonomous Ground Vehicle Group at IIT Kharagpur, under the mentorship of Professor Debashish Chakravarty.

I began my research journey with internships at institutions, including the ARMS Lab at IIT Bombay, where I worked on path planning and decision-making for connected autonomous vehicles under Professor Arpita Sinha, developing lane-change policies in CARLA simulations that comply with Responsibility-Sensitive Safety (RSS) rules. Prior to that, I was a research intern at the I3D Lab at IISc Bangalore, working on tactile manipulation and reinforcement learning under Professor Pradipta Biswas, focusing on designing RL pipelines and behavior cloning strategies for exploration tasks using tactile data.

Driven by a vision where robots assist humans seamlessly by performing complex physical tasks, I aspire to contribute impactful research in robotics that fosters collaborative and generalizable manipulation capabilities for real-world applications.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  CV

profile photo

Research

My research focuses on reinforcement learning, multi-agent systems, and robotic manipulation. I am particularly interested in developing robust learning algorithms that enable robots to adapt and collaborate in dynamic environments.

Currently, I am exploring tactile-based reinforcement learning for surface reconstruction and manipulation, as well as skill-based communication in multi-agent systems.

project image

Curvature Informed Furthest Point Sampling


Shubham Bhardwaj, Soumojit Bhattacharya*, Ashwin Vinod*, Aryan Koganti, Aditya Sai Ellendula, Balakrishna Reddy
arXiv 2024, 2024
arxiv /

Proposed CFPS, a curvature-informed point cloud sampling method using reinforcement learning to optimize point selection. Paper under review arXiv
project image

Blind Tactile Exploration for Surface Reconstruction


Yashaswi Sinha*, Soumojit Bhattacharya*, Yash Kumar Sahu and Pradipta Biswas
ICRA 2025(accepted), 2024

Developed a tactile exploration algorithm for reconstructing fine surface details of objects using a sequential controller-based approach. ICRA 2025
project image

DiffClone: Enhanced Behaviour Cloning in Robotics with Diffusion-Driven Policy Learning


Sabariswaran Mani, Sreyas Venkataraman*, Abhranil Chandra*, Adyan Rizvi*, Yash Sirvi*, Soumojit Bhattacharya*, Aritra Hazra
TOTO Benchmark|Neurips 2023, 2010
arxiv /

Developed DiffClone for offline RL tasks involving pouring and scooping with sparse rewards. arXiv

Other Projects

These include coursework, side projects, and unpublished research work.

project image

Scalable Multi-Agent Robot Swarm Navigation

Developed an end-to-end pipeline for multi-agent pathfinding in dynamic environments using reinforcement learning techniques.

project image

Skill-Based Communication for Multi-Agent Reinforcement Learning

Developed a skill-based communication framework for MARL to enhance coordination and scalability in partially observable environments.

project image

Student Satellite Program - Communication System Design

Designed and tested CubeSat communication systems, overcoming obstacles such as trees and buildings to establish successful ground-based transmission.


Design and source code from Jon Barron's website