Research

The AVIS Lab builds intelligent systems that perceive the world through cameras and sensors, reason about what they see, and act on it. Our work runs from foundations to deployment. On the foundations side we develop new learning algorithms, optimizers, and visual tokenizers. On the deployment side our systems fly drones, diagnose crop and human disease, and control assistive devices. We care equally about the science and the real-world result. Much of our work starts with a problem in Bangladesh, such as a field disease, a hospital workflow, or a sensing gap, and ends as a model, a dataset, or a working prototype.

Research areas

Robot learning & embodied intelligence

How should a robot decide how, when, and where to move? We study offline goal-conditioned reinforcement learning, where agents learn long-horizon, goal-reaching behaviour purely from logged data. Recent work includes GTRL, which grounds divide-and-conquer value learning with temporal differences, and a study showing that the agent’s state representation, not the goal representation, is the real bottleneck. This line builds on Dr. Hasan’s industrial experience with vision-based robot control (visual servoing, tool pick-up, and placement). It extends to compliant grasping and sensor-driven prosthetic hands.

Vision-language models & multimodal reasoning

Large vision-language models describe images well but often reason poorly about them. We build benchmarks and methods that test and improve multi-step, grounded visual reasoning. PlantInquiryVQA (Findings of ACL 2026) evaluates whether models can question a leaf image the way an expert botanist does. Related work covers counting objects hidden behind occluders, non-neural visual tokenization, and explainable vision-language systems for medical reports.

Autonomous UAVs

We want drones that can fly safely with nothing more than a single camera. VLM-Nav (PLOS One, 2026) turns monocular images into depth maps and lets a vision-language model plan collision-free manoeuvres without a prebuilt map. A follow-up benchmarks six VLMs in Unity–ROS 2 simulation and flight-tests them on a Pixhawk 6C quadrotor. Earlier work recognizes human actions from aerial video for drone surveillance (U-ActionNet, FFT-UAVNet).

Plant phenomics & precision agriculture

Plant phenotyping measures traits such as growth, architecture, yield, and stress tolerance, the quantitative basis of crop breeding. We use field imaging, UAVs, and deep learning to measure these traits automatically. Examples include detecting and segmenting wheat spikes (WheatSpikeNet), classifying cotton seedling growth stages from UAV imagery, and diagnosing plant diseases with lightweight on-device models (PlantCareNet, SugarcaneShuffleNet). Our current project, PlantPhenoBot, brings robotics into the field with Gazipur Agricultural University.

Medical imaging & biomedical AI

Clinicians need AI they can trust. We develop interpretable models for medical images and biosignals. This includes information-geometric networks for knee osteoporosis from X-rays and explainable AI for histological cancer detection (a MOST-funded project with Bangladesh Medical University). On the biosignal side we work on EEG decoding for stroke rehabilitation and on brain–computer interfaces for prosthetic control.

Intelligent sensing, IoT & learning foundations

We pair smart sensors with machine learning for environmental and industrial condition monitoring. Examples include IoT-driven energy forecasting for commercial buildings, landslide-risk prediction from climate projections, and real-time face recognition for video surveillance. Alongside these applications we work on the learning machinery itself: new optimizers (SPARK), loss functions (AGL Loss), and physics-informed machine learning for classifying celestial objects.

Highlighted publications

Selected recent work: new algorithms, a Findings of ACL paper, and articles in leading journals.

GTRL: Grounding Divide-and-Conquer Value Learning with Temporal Differences

Abdul Monaf Chowdhury, MD Sameer Iqbal Chowdhury, Shifat E Arman, Md Mehedi Hasan

arXiv  ·  27 Sep 2026  ·  arxiv:2609.33259

GTRL (Grounded Transitive RL) — an offline goal-conditioned value-learning algorithm that grounds divide-and-conquer updates with a one-step temporal-difference target, so every state–goal pair receives an update. Achieves the highest average success rate across nineteen OGBench tasks spanning stochastic, deterministic, and stitching environments.

arXiv

reinforcement-learning

Thinking Like a Botanist: Challenging Multimodal Language Models with Intent Driven Chain-of-Inquiry

Syed Nazmus Sakib, Nafiul Haque, Shahrear Bin Amin, Hasan Muhammad Abdullah, Md Mehedi Hasan, Mohammad Zabed Hossain, Shifat E. Arman

Findings of the Association for Computational Linguistics: ACL 2026  ·  22 Apr 2026  ·  doi:10.18653/v1/2026.findings-acl.1741

Findings of ACL 2026. Introduces PlantInquiryVQA, a benchmark of 24,950 expert-curated plant images and 138,068 question–answer pairs for multi-step, intent-driven visual reasoning in plant disease diagnosis.

ACL Anthology

arXiv

vision-language-models plant-phenomics

VLM-Nav: Mapless UAV navigation using monocular vision driven by vision-language models

Gobinda Chandra Sarker, AKM Azad, Sejuti Rahman, Md Mehedi Hasan

PLOS One  ·  03 Apr 2026  ·  doi:10.1371/journal.pone.0345778

Mapless obstacle avoidance for UAVs from a single monocular camera. RGB frames are converted to depth maps with DepthAnything-V2, and a vision-language model reasons over them to choose the next manoeuvre.

PLOS One

uav vision-language-models

PlantCareNet: an advanced system to recognize plant diseases with dual-mode recommendations for prevention

Muhaiminul Islam, AKM Azad, Shifat E. Arman, Salem A. Alyami, Md Mehedi Hasan

Plant Methods  ·  23 Apr 2025  ·  doi:10.1186/s13007-025-01366-9

An end-to-end plant-disease diagnosis system that pairs a CNN classifier with expert-guided, knowledge-based recommendations for prevention and treatment.

plant-phenomics precision-agriculture

Cotton seedling monitoring and growth stage classification integrating deep learning and feature engineering

Hasan Muhammad Abdullah, Muhaiminul Islam, Md. Shahinul Islam, Sudip Sen, Abdul Kaium Tuhin, Shifat E. Arman, Md Mehedi Hasan

Smart Agricultural Technology  ·  01 Dec 2025  ·  doi:10.1016/j.atech.2025.101619

A high-throughput phenotyping framework combining SeedlingNet, a lightweight residual CNN trained on UAV and ground-level imagery, with a biologically grounded Composite Stress Index to classify cotton growth stages and flag at-risk seedlings early.

plant-phenomics precision-agriculture

Development of an IoT-driven deep learning framework for smart energy prediction in commercial buildings

Gobinda Chandra Sarker, Md Mehedi Hasan, Md. Rakibul Hoque, Md. Khashrul Alam, M. Helal Uddin Ahmed

Journal of Building Engineering  ·  01 Oct 2025  ·  doi:10.1016/j.jobe.2025.113315

An IoT-driven deep learning framework for predicting energy consumption in commercial buildings.

iot deep-learning

All publications

Every paper from 2022 onward, grouped by year and searchable, is on the Publications page.