About me

Chulabhaya Wijesundara

I’m CK, a PhD student at Northeastern University and a research engineer at STR.

My research focuses on multi-agent reinforcement learning, learning under partial observability, and policy optimization. At Northeastern, I work with Christopher Amato in the Lab for Learning and Planning in Robotics. At STR, I work on applied research projects spanning reinforcement learning and computer vision.

Outside of work, I enjoy lifting weights, playing video games, reading, photography, and improving my iced latte game.

Read my papers My work at STR Google Scholar

01 / Publications

Selected papers

All papers

NeurIPS 202601

Rethinking Ratio-Based Trust Regions for Policy Optimization in Multi-Agent Reinforcement Learning

Rethinking policy updates to make cooperative multi-agent reinforcement learning more stable.

Project page Paper

TMLR 202502

Leveraging Fully-Observable Solutions for Improved Partially-Observable Offline Reinforcement Learning

Using state information during training to improve offline reinforcement learning with incomplete observations.

Project page Paper

02 / Work

Work at STR

Here are a few highlights from my work at STR, where I apply reinforcement learning to a variety of problems.

Algorithm leadership

DARPA AIR

I led the algorithm development that earned STR a Phase 2 award in DARPA’s Artificial Intelligence Reinforcements (AIR) program.

About DARPA AIR

Simulation transfer

Zero-shot transfer

I developed and demonstrated techniques for transferring RL agents trained in lower-fidelity simulations to higher-fidelity models of real-world systems without additional training.

From PhD to programs

JAX-based multi-agent RL

I transitioned JAX-based multi-agent RL research and software infrastructure from my PhD into active STR programs. This work now supports a range of programs across the company.

Adversarial learning

Human surrogates and visual attacks

I developed RL-based human surrogate agents for mixed-reality environments using VLM perception, alongside adversarial agents trained to discover novel visual attacks and expose weaknesses in surrogate perception and robustness

03 / Software

Tools and benchmarks

Frameworks and environments for training and evaluating RL agents.

Training framework

MAPX

A JAX framework for multi-agent policy-gradient reinforcement learning research.

JAXMulti-agent RLRecurrent policies

In development

Simulation & benchmarks

AeroJAX

JAX-native aerial environments for studying decentralized control and coordination in continuous three-dimensional space.

JAXContinuous control3D environments