Chandra (pronunciation: chun-dra; chun rhymes with sun)
I'm a Masters student in Computer Science at Brown University working with George Konidaris and Chen Sun. My research goal is to build robots that can learn from existing experience and continue improving safely when they encounter unfamiliar tasks and environments. To this end, I am interested in Offline RL, Offline-to-Online RL and Safe RL.
Previously, I graduated with a Bachelors in Computer Science and Engineering from IIT Tirupati, where I worked with Ajin George Joseph on Risk-Sensitive RL. During this time, I was fortunate to spend a year as a Student Researcher at Robert Bosch Center for Data Science and AI working with Balaraman Ravindran on Explainable AI for Graphs to ground LLM reasoning.
I'm actively looking for and applying to PhD positions starting Fall 2027!
Outside of research, I enjoy playing sports (mainly cricket and fútbol, although I'd never say no to any sport!), trekking, cooking and editing videos!
Google Scholar / Twitter / LinkedIn / GitHub / CV
Email: chandradithya_jonnalagadda[at]brown[dot]edu
Presented our work on Hierarchical RL with Temporally-extended Latent Action World Models and recieved the Best Paper Award at the RLC 2026 Model-Based RL in the Era of Generative World Models Workshop in Montreal!
Our work titled "Latent Actions from Factorized Transition Effects under Agent Ambiguity" was accepted to the ICML 2026 Compositional Learning Workshop!
Our work titled "The Price of Paranoia: Robust Risk-Sensitive Cooperation in Non-Stationary Multi-Agent Reinforcement Learning" was accepted to the AAMAS 2026 Adaptive Learning Agents Workshop!
Started an ScM in Computer Science at Brown University!
Graduated from IIT Tirupati with a BTech in Computer Science and Engineering!
Thesis: Beyond Expectation: Tail-Sensitive RL through Exponential Utility.
RLC 2026 · Model-Based RL in the Era of Generative World Models Workshop · Best Paper Award 🏆
We introduce an offline goal-conditioned framework that learns a high-level world model used to plan subgoals for a value-based model-free RL policy. We propose Temporally-extended Latent Action World Models (Te-LAWMs) as high-level world models, trained via a latent action modeling objective.
ICML 2026 · Compositional Learning Workshop
paper / project page / code
We propose Observed Transition Factorization LAM (OTF-LAM), a latent action model that learns environment-agnostic motion primitives from visual changes and composes them into state-aware latent actions that capture coordinated and independent movements. These latent actions are mapped to real actions using a small amount of action-labeled data.
AAMAS 2026 · Adaptive Learning Agents Workshop
We propose a MARL algorithm that sustains cooperation between agents in mixed-motive games and social dilemmas by improving robustness to the policy gradient update variance induced by the other agents' uncertainties.