Rocktim Jyoti Das

I am a second-year Computer Science PhD student at University of Maryland, College Park, working with Prof. Dinesh Manocha. My research is on physically grounded robotics. I work on physics-based world models that let robots predict how objects deform, move, and respond to contact and force input. I also design tactile sensors that give robots the touch feedback they need to manipulate the physical world reliably.

Before starting my PhD, I was a Research Associate II at MBZUAI in Abu Dhabi, working on robotic manipulation and reasoning under the supervision of Prof. Ivan Laptev and Prof. Preslav Nakov. I developed intelligent agents that reason, plan, and interact with the environment to solve complex tasks, using world knowledge learned from internet-scale data.

I completed my undergraduate studies at the Indian Institute of Technology, Delhi, which set the foundation for my academic journey. I had the incredible opportunity to spend a year at the DAIR Lab, IIT Delhi, as a Research Assistant under the supervision of Prof. Mausam. At DAIR Lab, I focused on task-oriented dialog systems and conversation-based medical diagnosis.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github /  LinkedIn /  Reading Group

News

• 08/2024 Preprint of BLAZER: Bootstrapping LLM-based Manipulation Agents with Zero-Shot Data Generation released.
• 07/2025 Started PhD at University of Maryland, College Park under the supervision of Prof. Dinesh Manocha.
• 06/2025 Our paper MALMM: Multi-Agent Large Language Models for Zero-Shot Robotics Manipulation is accepted as an oral presentation at IROS, 2025.
• 11/2024 Attending EMNLP, 2024 in Miami, Florida.
• 10/2024: The LLaMA-3-Nanda-10B-Chat model is released, establishing itself as the leading open-sourced bilingual Hindi-English model.
• 09/2024 3 papers accepted in EMNLP, 2024.
• 09/2024 Joined Embodied Perception Group at MBZUAI under the supervision of Prof. Ivan Laptev.
• 08/2024 Attending ACL, 2024 in Thailand.
• 05/2024 Our paper EXAMS-V is accepted in ACL main, 2024.
03/2024 Preprint of EXAMS-V: A Multi-Discipline Multilingual Multimodal Exam Benchmark for Evaluating Vision Language Models released.
12/2023 Attending EMNLP, 2023 in Singapore.
06/2023 Joined Mohamed bin Zayed University of Artificial Intelligence as a Research Associate under the supervision of Prof. Preslav Nakov.
05/2023 Our paper Exploring distributional shifts in large language models for code analysis accepted in EMNLP, 2023.
05/2023 Our paper DKAF: KB Arbitration for Learning Task-Oriented Dialog Systems with Dialog-KB Inconsistencies accepted in ACL Findings, 2023.
11/2022 Awarded Rajiv Bambawale Cash Prize Award for best undergraduate thesis.
11/2022 Graduated from Indian Institute of Technology Delhi with a Bachelors in Electrical Engineering and minor in Computer Science.
10/2022 Started remote internship at Intelligence and Knowledge Discovery Research Lab, USC under the supervision of Prof. Xiang Ren.
06/2022 Joined Data Analytics and Intelligence Research (DAIR) Group, IIT Delhi as a Project Scientist under the supervision of Prof. Mausam.

Selected Research

* denotes joint first authors

Open-Sourced Project

Nanda is the leading open-sourced Hindi bilingual model for generation and reasoning, surpassing current models on safety benchmarks. Although our model is optimized for Devanagari Hindi, it also seamlessly supports Romanized and code-mixed Hindi, making it highly versatile for real-world applications. Check out the paper for more details about training, evaluation!

📄 Paper link

Give our model a try!

Download LLaMA-3-Nanda-10B weights from Hugging Face and see its capabilities in action:

💾 Model weights

Reading Group

I organize the UMD Robotics Reading Group, a weekly meeting (Tuesdays, 5–6 pm) where we read and discuss recent papers on robot learning, manipulation, and humanoid robots. See the schedule for upcoming sessions and slides.

Source code from Jon Barron's website.