Inference-Time Alignment of Diffusion-Based Trajectory Planners for Autonomous Driving

Prof. Sotirios Chatzis, Cyprus University of Technology

When: Wednesday, 1st July 2026, at 10:00 AM
Where: IMBIT, NEXUS Lab, Georges Köhler Allee 201, 79110 Freiburg

Abstract: Diffusion models have emerged as a powerful class of generative trajectory planners for autonomous driving, capable of producing diverse, multimodal motion plans from large-scale demonstration data. Yet deploying such models in safety-critical settings raises a recurring challenge: how to steer pretrained diffusion planners towards specific downstream objectives — task-level rewards, safety constraints, or stylistic preferences — without compromising the rich behavioural priors learned during pretraining. The dominant approach is reinforcement-learning fine-tuning of the underlying policy, which is effective but costly, can degrade the generative prior, and is non-trivial to certify.

In this talk he will present recent directions from our group on inference-time alignment of diffusion-based planners, where alignment with task objectives is performed at sampling time over the trajectory space of a fixed, pretrained planner, without modifying its weights. He will motivate the formulation, walk through the key design choices they have been exploring, and share preliminary closed-loop observations on standard driving benchmarks. He will close by relating the approach to broader questions of trustworthy and verifiable autonomy that they are pursuing in the context of the EU HIDDEN project.

Bio: Sotirios Chatzis is a Full Professor and Department Chair in the Department of Electrical Engineering, Computer Engineering and Informatics at the Cyprus University of Technology, where he directs a research lab on machine learning. His work lies at the intersection of Bayesian deep learning, scalable generative modelling, and trustworthy AI, with applications to autonomous systems, multimodal foundation models, recommendation systems, and decision-making under uncertainty. He has published extensively at top venues including NeurIPS, ICML, ICCV, AAAI, and AISTATS, and has served as Principal Investigator on numerous European Commission and Cyprus Research and Innovation Foundation projects. He is a member of the National AI Taskforce of the Republic of Cyprus. Prior to joining CUT, he was a Senior Research Associate at Imperial College London. He holds a degree in Computer Engineering and a PhD in Statistical Machine Learning.

Host: Prof. Abhinav Valada, Robot Learning Lab