PhD researcher at UMBC · MPSC Lab · CYPRESS · CARDS (ArtIAMAS program)
Robot learning for reliable autonomy.
Research focuses on reinforcement learning and embodied AI under uncertainty and physical constraints. Learned representations of motion cost, risk, and recoverability guide decisions in navigation and whole-body control.
Current work studies how to use demonstrations selectively for humanoid locomotion and manipulation, manage risk in language-guided navigation, and transfer policies from simulation to physical robots.
Research areas
I am on the job market and seeking full-time research opportunities in industry and academia. I also welcome collaborations in robot learning and embodied AI.
Updates & milestones
Current research explores language-guided navigation and narrow-gap traversal under uncertainty. Learned risk estimates guide route selection and recovery, while control based on available clearance adjusts body alignment through tight passages. The goal is to connect task understanding with the physical constraints that determine whether a route can be completed reliably.
Current work studies where demonstration guidance matters most for whole-body humanoid learning. Physical margins, required contacts, and predicted near-term failure determine how closely the policy should follow a demonstrated action. Guidance increases in restrictive states and relaxes where reinforcement learning can explore effective alternatives, across tasks such as stairs, object pickup, sitting, and carrying.
10 publications
CoRL 2026
CoGeL is a safety-aware long-horizon navigation framework that learns from demonstrations, failures, and interventions. It turns local safety evidence into a directed navigation distance, allowing robots to avoid hidden hazards and plan safer routes over long horizons.
ICML 2026
This paper introduces a Finslerian reinforcement learning framework for risk-aware locomotion in environments with asymmetric and anisotropic traversal costs. Instead of treating the cost between two states as a symmetric distance, FiRL models motion using a direction-dependent geometry, where the effort of moving through the environment depends on both the current state and the direction of travel. This is captured through a Finsler-style cost function F(s,s'), enabling the learned policy to represent behaviors and direction-sensitive locomotion risk. We further incorporate tail-risk sensitivity so that the policy does not optimize only for average performance, but also accounts for rare, high-cost failures that are critical in real-world robot deployment.
ICCCN 2026
This paper introduces a simulation-enhanced approach for realistic navigation planning in multi-agent robotic systems, optimizing strategies for operations in contested and unpredictable environments.
ICML 2025
In this work, we address the challenge of modeling asymmetric costs in goal-reaching tasks, where the effort to move between two states can differ depending on direction. We introduce a decomposition of the quasimetric function: d(s, g) = Φ(g) − Φ(s) + Ψ(s → g). The term Φ(g) − Φ(s) captures the path-independent potential difference, representing the global structure of the environment, while Ψ(s → g) encodes the path-dependent residual that models asymmetries and directional effects. This formulation enhances sample efficiency by structuring the learning problem and incorporates a Lyapunov-based safety mechanism that ensures stable and reliable exploration.
ICRA 2025
This paper presents an asymmetric cost-aware navigation method using quasimetric-constrained policy optimization to improve navigation planning in complex, real-world environments with uneven costs and constraints.
IROS 2024
A navigation framework that leverages hierarchical reinforcement learning (HRL) to enable autonomous robots to efficiently explore unknown areas. By building active topological maps and incorporating intrinsic rewards, TopoNav enhances navigation accuracy and adaptability, especially in sparse-reward environments.
ACSOS 2024
This approach enhances maneuver planning by using target-aware visibility estimations and offline reinforcement learning, making it highly suitable for complex and contested environments where maintaining cover is critical.
ACSOS 2023
CoverNav presents a deep reinforcement learning approach for navigation planning that prioritizes cover following in unstructured outdoor environments.
IEEE DCOSS 2022
SynchroSim provides an integrated co-simulation middleware that supports coordination and interaction among heterogeneous multi-robot systems. It ensures effective communication and synchronization across diverse platforms, allowing robust development and testing of cooperative robotic applications in dynamic environments.
Under review
GapNav is a margin-aware navigation framework that enables legged robots to enter and traverse narrow passages near their physical clearance limits. It estimates the available side margin and adjusts the robot’s motion, orientation, and body height to reduce collisions throughout the crossing.
Research & background
Center for Research in Use-Inspired Cyber-Physical Systems (CYPRESS)
University of Maryland, Baltimore County
Baltimore, MD
Aug 2025 – Present
Center for Distributed Sensing and Autonomy (CARDS)
University of Maryland, Baltimore County
Baltimore, MD
Aug 2022 – Aug 2025
Stormfish Scientific Corporation
Baltimore, MD
May 2024 – Aug 2024
Developed reinforcement learning models for navigation and connected ROS-based robots with AuroraXR and Unity simulations.
Mobile, Pervasive and Sensor Computing Lab (MPSC)
University of Maryland, Baltimore County
Baltimore, MD
Aug 2021 – Present
BJIT Limited
Dhaka, Bangladesh
Oct 2018 – Aug 2021
Worked on the PockeTalk translation SDK, image translation integration, and the JAJA TV application.
IPvision Canada Inc.
Dhaka, Bangladesh
Sept 2015 – Sept 2018
Developed distributed cloud storage systems using OpenStack Swift.
Eyeball Networks Inc.
Dhaka, Bangladesh
July 2014 – Aug 2015
Developed the Android uGrow Smart Baby Monitor application for Philips Avent hardware.
University of Maryland, Baltimore County
2021 - present
Advisor: Prof. Nirmalya Roy. Research in reinforcement learning, robot navigation, and reliable autonomy.
University of Maryland, Baltimore County
2021 - 2023
Shahjalal University of Science and Technology
2009 - 2013
Specialization in machine learning.
I am on the job market and seeking full-time research opportunities in industry and academia. I also welcome collaborations in robot learning and embodied AI.