Robust Policy Gradient Optimization through Action Parameter Perturbation in Reinforcement Learning
ARLET Workshop at NeurIPS 2025
Md Masudur Rahman
Postdoctoral Researcher, Purdue University
I am a postdoctoral researcher in the Edwardson School of Industrial Engineering at Purdue University, working with Juan P. Wachs. I received my Ph.D. in Computer Science from Purdue in 2024, advised by Yexiang Xue.
My research is in reinforcement learning and robot learning, with a focus on surgical and medical robotics. I work on policy optimization methods that remain reliable when conditions change, on vision-language models that support clinical decisions, and on semi-autonomous systems in which a surgeon and a robot share control. The common goal is co-adaptive intelligence: AI that learns through interaction with living systems.
My Robust Policy Optimization (RPO) method is included in the CleanRL and skrl reinforcement learning libraries and was used to train simulation expert policies in research on real-time action chunking for robot policies. My work on AI-assisted burn diagnosis has been published in Annals of Surgery, presented as an oral at AAAI/IAAI 2026, and recognized with the Best Paper Award at MIUA 2025.
rahman64@purdue.edu · CV (PDF) · Google Scholar · GitHub · LinkedIn
News
All newsARISE, on agentic vision-language reasoning for interpretable IBD assessment, was accepted at AAAI 2027 (IAAI-27). The work is led by Pronoma Banerjee, a Ph.D. student I mentor.
Received a NAIRR compute allocation as principal investigator, on Argonne’s Cerebras CS-3 system.
“A Multimodal Artificial Intelligence Reasoning Framework for Burn Diagnosis” was published in Military Medicine.
Recognized as a Gold Reviewer at ICML 2026.
AMBUSH-AI, our system for automated noninvasive burn diagnosis, was accepted at Annals of Surgery.
AURA was accepted for an oral presentation at IAAI 2026.
Received the Best Paper Award at MIUA 2025 for work on knowledge-driven burn diagnosis with vision-language models (Purdue news).
Selected publications
All publications
Automated Unified Reasoning with Vision-Language Models for Multi-modal Burn Assessment
AAAI 2026 (IAAI-26). Oral presentation
Annals of Surgery, 284(2), 2026
Knowledge-Driven Hypothesis Generation for Burn Diagnosis from Ultrasound with Vision-Language Model
MIUA 2025. Best Paper Award
Natural Language-based State Representation in Deep Reinforcement Learning
Findings of NAACL 2024
SARTRES: A Semi-Autonomous Robot Teleoperation Environment for Surgery
Computer Methods in Biomechanics and Biomedical Engineering: Imaging & Visualization, 9(4), 2021
Funding and awards
Full record in CV- NSF EAGER, “Theoretical Foundations for Integrating Foundational Models into Reinforcement Learning,” 2025–2027. Key personnel (PI: Juan P. Wachs).
- NAIRR compute allocation on the Cerebras CS-3 system at Argonne National Laboratory, 2026. Principal investigator.
- Pilot funding from the Purdue Health of the Forces program and the Clinical and Translational Sciences Institute, 2025–2027. Team lead (PI: Juan P. Wachs).
- Best Paper Award, MIUA 2025.
- Oral presentation, AAAI/IAAI 2026.
- Gold Reviewer, ICML 2026.
Teaching and mentoring
DetailsI have given guest lectures in Purdue’s IE 549: Machine Vision in Intelligent Robotic Systems and am co-teaching a course on AI in medicine and large language models. I currently mentor three Ph.D. students in Industrial and Biomedical Engineering on medical AI, reinforcement learning with foundation models, and medical robotics.