Academic Work

Research & Publications

Undergraduate Researcher · Cognitive Assistive Robotics Lab · UNH · Advisor: Prof. Momotaz Begum

Safe and reliable control of long-horizon manipulation by learned visuomotor policies, built on latent dynamics that carry stability guarantees.

Visuomotor PoliciesKoopman OperatorsFlow MatchingSkill DiscoveryImitation LearningInterpretability

$ robots --hardware Franka FR3 · Boston Dynamics Spot · Hello Robot Stretch 3/4 · ABB YuMi · Inspire Dexterous Hand

<Research Themes/>

01

Reliable generative policies

Making flow- and diffusion-based visuomotor policies precise on real robots, and collecting the demonstrations they learn from.

02

Structured dynamics & skills

Latent dynamics with stability guarantees, and skill boundaries that a planner can reason about.

<Publications/>

IROS 2026

Trajectory-Consistent Flow Matching for Robust Visuomotor Policy Learning

Riad Ahmed, Sujosh Nag, Moniruzzaman Akash, Mostafa Hussein, Momotaz Begum

Flow-matching policies optimize a pointwise velocity objective but integrate that field at inference, and the mismatch compounds along the trajectory. The paper closes it with auxiliary rectified-flow velocity regression, multi-step consistency training on the integrated displacement of trajectory segments, velocity-field regularization for temporal smoothness, and Runge-Kutta integration in place of Euler; the ablation shows none of the four suffices alone.

Flow MatchingFranka FR3SpotPoint Clouds
arXiv ↗

<Research Projects/>

Independent research

Structured Belief-Guided Diffusion Policy

Conditioned a diffusion policy on a belief over task phase and progress predicted from image observations, taking as given the skill boundaries that define those phases. The belief helped in Meta-World and on real Franka tasks, yet under temporal perturbation the policy recovered less reliably than an unconditioned diffusion policy. That negative result moved my attention upstream, to where skill boundaries should come from at all.

Diffusion PolicyMeta-WorldFranka FR3

Independent research

Belief-PDDL: Neuro-Symbolic Planning Under Partial Observability

Planning framework that maintains factorized Bayesian beliefs over grounded predicates, projects them through CP-SAT constraints to drop inconsistent states, and plans over Top-K feasible worlds with active sensing; robust to heavy observation noise in Blocksworld and ALFWorld.

PDDLCP-SATALFWorld
⌥ Code

<Course Research/>

Multi-Agent Grid Environment

A
B
Agent A
Agent B
Wall

AI course

Semantic Observation Repair for Decentralized Multi-Agent Path Finding

Inference-time repair of corrupted observations for a frozen decentralized PPO policy in POGEMA, across eight corruption types (noise, dropout, bit flips, FGSM/PGD). Over 4,000 episodes, semantic projection to valid observation grids raised mean success from 68.8% to 88.2% and recovered FGSM success from 0% to 97.5% without retraining.

MAPFPOGEMAAdversarial Robustness
⌥ Code

ML course

AAPR: Adversary-Adaptive Privacy-Preserving Representations

Suppressed biometric attributes (speaker identity, gender) in learned representations while keeping task utility, via a variational information bottleneck filter trained through a gradient reversal layer against a periodically re-initialized adversary. Gender leakage fell to chance but identity stayed fully recoverable.

PrivacyInformation BottleneckAdversarial Training
⌥ Code