Prajwal Koirala

I am a Ph.D. student in Robotics at Cornell University, advised by Professor Mark Campbell. My research focuses on learning and estimation for robotics, with an emphasis on developing intelligent, reliable robotic systems that operate safely in uncertain and dynamic environments. I am particularly interested in integrating machine learning with principled models for perception, decision-making, and control to improve robustness, safety, and real-world generalization.

Before joining Cornell, I received my MS from Iowa State University, where I worked on reinforcement learning and safety-critical control for autonomous systems under the guidance of Dr. Cody Fleming and in frequent collaboration with Dr. Soumik Sarkar. My research included developing offline reinforcement learning algorithms that integrate safety constraints into data-driven decision-making, resulting in publications at leading conferences in machine learning, robotics, and controls (ICLR, CoRL, CDC, among others). I also gained teaching experience as a Graduate Teaching Assistant, conducting lab sessions in system dynamics and control and received multiple Research Excellence Awards for my contributions during my time at Iowa State.

News

Jun 17, 2026	Two papers accepted to IROS 2026. One presents a decentralized generative framework for multi-agent coordination, where distilled communication enables effective decision-making under partial observability. The other develops a value-guided flow-matching approach for robot policy learning, enabling scalable integration of RL objectives into expressive controllers. Preprints and code to be released soon!
May 15, 2026	LexiSafe received a Best Paper Nomination at IEEE/ACM ICCPS 2026. The work introduces a lexicographic prioritization framework for safe RL using partial policy freezing to reduce safety violations while maintaining strong performance in safety-critical CPSs, with an extension to multiple hierarchical safety objectives.
Jan 26, 2026	Flow-Based Single-Step Completion for Efficient and Expressive Policy Learning has been accepted to ICLR 2026. We propose a flow-based generative policy for one-shot action generation, enabling much faster training and inference (across offline RL, GCRL, and BC).
Aug 22, 2025	Thrilled to begin my Ph.D. in Robotics at Cornell University! Excited to explore cutting-edge research in safe and intelligent autonomy. Supported in my first semester by the Cornell Fellowship.
Jul 15, 2025	Paper accepted to CDC 2025, Rio de Janerio. “FAWAC: Feasibility Informed Advantage Weighted Regression for Persistent Safety in Offline Reinforcement Learning” Coauthored with Dr. Zhanhong Jiang, Dr. Soumik Sarkar and Dr. Cody Fleming.

Selected Publications

Find a complete list of the publications in my Google Scholar profile .

ICLR
2025

Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning

Prajwal Koirala, Zhanhong Jiang, Soumik Sarkar, and 1 more author

arXiv preprint arXiv:2412.08794, 2024

arXiv Bib

@article{koirala2024latent,
  title = {Latent Safety-Constrained Policy Approach for Safe Offline Reinforcement Learning},
  author = {Koirala, Prajwal and Jiang, Zhanhong and Sarkar, Soumik and Fleming, Cody},
  journal = {arXiv preprint arXiv:2412.08794},
  year = {2024},
}

CoRL
2024

Solving offline reinforcement learning with decision tree regression

Prajwal Koirala and Cody Fleming

In 8th Conference on Robot Learning, 2024

arXiv Bib

@inproceedings{koirala2024solving,
  title = {Solving offline reinforcement learning with decision tree regression},
  author = {Koirala, Prajwal and Fleming, Cody},
  booktitle = {8th Conference on Robot Learning},
  year = {2024},
  organization = {PMLR},
}

ICLR
2026

Flow-Based Single-Step Completion for Efficient and Expressive Policy Learning

Prajwal Koirala and Cody Fleming

arXiv preprint arXiv:2506.21427, 2025

arXiv Bib

@article{koirala2025flow,
  title = {Flow-Based Single-Step Completion for Efficient and Expressive Policy Learning},
  author = {Koirala, Prajwal and Fleming, Cody},
  journal = {arXiv preprint arXiv:2506.21427},
  year = {2025},
}