CV

Postdoctoral Researcher, Representation Learning, Multi-Modality and Generalization

VITA Lab, EPFL | Valeo AI

Build SOTA transformer‑based frameworks for multi-modal motion prediction using representation learning; self-distillation and masked reconstruction and develop cross-domain generalization techniques, parameter efficient fine-tuning strategies, and robustness measures.

Postdoctoral Researcher, Trajectory Prediction and Planning

ASTRA Lab, INRIA

Integrated spatio-temporal transformer models for trajectory prediction and planning in dynamic environments.

Visiting Researcher, Context-based Trajectory Prediction for Autonomous Driving

LISA Lab, UCSD (USA)

Proposed an approach using Multi-Head-Attention (MHA) and a joint representation of the static scene latent features and surrounding agents and Achieved second place in the nuScenes trajectory prediction challenge with the MHA-JAM model.

Ph.D. Researcher, Attention-based Trajectory Prediction for Autonomous Driving

RITS Lab, INRIA | Sorbonne University

Explored fundamental challenges in deep learning-based motion prediction and interaction dynamics within complex environments, emphasizing multi-modal generation and physical feasibility.

PhD in computer science/vision

Sorbonne University - INRIA Paris

Deep Learning for Motion Prediction

Engineering Studies

École Polytechnique de Tunisie, Top engineering school in Tunisia

Computer Science and Electrical Engineering

Preparatory Classes

Sfax Preparatory Engineering Institute

Fundamental Mathematics and Physics

George N. Saridis Best Transactions Paper Award of The IEEE TIV

2nd Position Award, ICRA 2020 Prediction Challenge

PerReg+: Generalizable Trajectory Prediction using Representation Learning

Accepted for CVPR 2025

Designed and implemented PerReg+, an advanced trajectory prediction model integrating Dual-Level Representation Learning with Self-Distillation and Masked Reconstruction to capture global context and granular details.

Forecast-PEFT: Efficient Fine-Tuning for Motion Prediction

Under Review

Designed Forecast-PEFT, a fine-tuning strategy significantly reducing training costs by freezing most model parameters and leveraging prompts and adapters. Enhanced prediction accuracy by up to 9.6% compared to conventional methods, achieving state-of-the-art efficiency with only 17% of the trainable parameters.

UniTraj: Unified Framework for Cross-Dataset Prediction

ECCV 2024

UniTraj: a framework that harmonizes diverse datasets, models, and evaluation metrics. It benchmarked various prediction models across multiple autonomous vehicle datasets, providing critical performance insights on prediction models’ cross-dataset generalizability and dataset characteristics and achieved state-of-the-art results on the nuScenes dataset by enhancing data diversity and scalability.

Social-Transmotion: Multi-Modal Human Motion Prediction

ICLR 2024

Designed Social-Transmotion, a transformer-based model that integrates diverse multimodal visual cues and employed masking techniques to ensure robustness against missing cues and validated on multiple datasets.

Manipulating Trajectory Prediction with Backdoors

Under Review

Identified vulnerabilities in motion prediction models to backdoor attacks and developed a defense mechanism using social attention to detect and mitigate abnormal agent influence.