Reinforcement Learning
Learning to reach any goal from a fixed dataset. I study value learning that stays correct over long horizons and under stochastic dynamics, and which representations actually help a goal-reaching agent act better.
Research · AVIS Lab, University of Dhaka
Abdul Monaf Chowdhury
Applying for PhD positions · Fall 2027I work on embodied agents that learn long-horizon skills from offline data and from language feedback.
I’m a Research Assistant at the AVIS Lab, University of Dhaka, working on offline goal-conditioned reinforcement learning and long-horizon robot manipulation. Most recently, I grounded divide-and-conquer value learning with temporal differences so it holds up under stochastic dynamics (GTRL), and turned a vision–language model’s reflections on failed episodes into reward for robot manipulation (LAGEA, ICML 2026).
Before this, I was a Research Assistant at the MAIM Lab, building a wearable fetal-movement monitor for stillbirth prevention, a Wellcome Leap In Utero project with Dr. Abhishek Kumar Ghosh and Dr. Niamh Nowlan. I hold a BSc in Robotics and Mechatronics Engineering from the University of Dhaka (2024), where Dr. Md Mehedi Hasan supervised my thesis on UAV-based human action recognition.
Research
I want agents that can carry out long, multi-step tasks in the real world. That takes value estimates that stay accurate over long horizons, feedback that tells an agent where an attempt went wrong and not just that it failed, and representations that bring together what an agent sees, reads and measures. My work so far follows three threads.
Embodied AI
Vision–language models as critics for robot learning: structured reflections on failed episodes, grounded in time and turned into dense reward shaping for manipulation.
Multimodal learning
Fusing complementary views of one signal, such as time, frequency and language, or the visible and hidden parts of a scene, into representations that hold up with little data.
PhD
Fall 2027
Looking for a PhD position
I’m applying to PhD programs starting in Fall 2027. I want to keep building embodied agents that learn long-horizon skills, where reinforcement learning, vision–language models and robot manipulation meet. My first-author work has appeared at ICML 2026 and AAAI 2026.
If you work on robot learning, embodied AI or reinforcement learning and think we’d be a good fit, I’d love to hear from you.
What I want to work on
- Reinforcement Learning
- Decision-making that scales to long horizons, whether learned offline from data or online through interaction.
- Embodied AI/Robot Learning
- Teaching robots new skills from demonstrations, data and their own experience.
- Multimodal Learning
- Models that understand vision, language and other modalitiess for perception and reasoning.
News
- Our paper
Beyond Visible Evidence: Open-World Amodal Counting Under Spatial Occlusiongot accepted toWACV '27 - Submitted
Contingency Planning with Latent World Models under Irreversible OutcomestoCVPR '27 - Submitted our journal paper from the Wellcome Leap In Utero project,
A Noise-Tolerant Wearable Device for Unsupervised In-Home Monitoring of Fetal Movements, toInformation Fusion - Submitted two papers to
ICLR '27:GTRL: Grounding Divide-and-Conquer Value Learning with Temporal DifferencesandDo Better Goal Representations Improve Goal-Conditioned Reinforcement Learning? - My paper titled
LAGEA: Language Guided Embodied Agents for Robotic Manipulationgot accepted for a Poster atICML '26 - My paper titled
T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusiongot accepted for a Poster atAAAI '26 - My journal paper titled
U-ActionNet: Dual-Pathway Fourier Networks with Region-of-Interest Module for Efficient Action Recognition in UAV Surveillance, based on my fourth year thesis work, has been published in IEEE Access - Our work on stillbirth prevention got another year of funding extension worth $1M by Wellcome Leap - In Utero
- Graduated from
University of Dhaka, Bangladeshwith a BSc. in Robotics & Mechatronics Engineering. [Report] - My paper titled
FFT-UAVNet: FFT Based Human Action Recognition for Drone Surveillance Systemgot accepted to the 5th IEEE International Conference on Sustainable Technologies for Industry 5.0 (STI) conference
Selected publications
-
arXiv 2026
GTRL: Grounding Divide-and-Conquer Value Learning with Temporal Differences
Grounding divide-and-conquer value learning with temporal differences, so offline goal-conditioned RL stays correct when the dynamics are stochastic.
-
Preprint 2026
Do Better Goal Representations Improve Goal-Conditioned Reinforcement Learning?
Handed a perfect goal representation, a goal-conditioned agent barely acts better. The headroom is in how it sees its own position.
Project Paper & code soon
-
ICML 2026
LAGEA: Language Guided Embodied Agents for Robotic Manipulation
A vision–language model reflects on each failed episode in structured language, and the reflection becomes time-localised reward shaping for robot manipulation.
-
AAAI 2026
T3Time: Tri-Modal Time Series Forecasting via Adaptive Multi-Head Alignment and Residual Fusion
T3Time reads a series three ways (in time, in frequency, and as a language prompt) and lets the forecast horizon decide how to weigh them.
-
WACV 2027
Counting Through Occlusion: Framework for Open World Amodal Counting
CountOCC counts the objects you cannot see, by reconstructing their features under the occluder and checking its attention against an unoccluded view.
Global visitors
A live snapshot of where visitors have been dropping in from. Click anywhere on the map to open the more detailed view.