|
I am a Master's student currently visiting MIT LIDS, advised by Navid Azizan. I am in the final semester of my Master's degree in Mechanical Engineering and Data Science at EPFL, where I worked with Caglar Gulcehre, Mathieu Salzmann, and Jürg Schiffmann. I did my Bachelor's degree in Aerospace Engineering at Politecnico di Milano. My current interests are in test-time guidance and fine-tuning of diffusion and flow matching models, stochastic optimal control, and applications to robotic planning and safe control. |
|
arXiv |
abstract |
code
Diffusion models have emerged as powerful tools for planning and control by learning multimodal distributions over actions and trajectories. Yet reliable inference-time safety enforcement remains a key barrier to their deployment in safety-critical tasks. Existing approaches typically project each denoising iterate onto the feasible set, even though constraints are defined only on the final clean trajectory. Enforcing feasibility on noisy intermediate samples can therefore overconstrain the sampling dynamics, substantially degrading sample quality. To address this limitation, we introduce DiRecT (Diffusion-based planning via Receding-horizon denoising with Terminal constraints), a training-free algorithm for constrained sampling from diffusion models via stochastic optimal control (SOC). DiRecT enforces constraints only on the final clean sample, avoiding unnecessary restrictions on the intermediate denoising dynamics. Inspired by model predictive control, we derive a principled receding-horizon surrogate for the otherwise intractable constrained SOC formulation, yielding an efficient algorithm that cleanly separates stochastic denoising from constraint satisfaction, progressively steering samples toward feasible final trajectories without distorting the learned diffusion dynamics. Furthermore, DiRecT is highly flexible: it can leverage off-the-shelf or domain-specific optimizers, incorporate priors over environment dynamics, and optimize additional soft rewards. Extensive experiments on safe planning benchmarks demonstrate that DiRecT substantially improves deployment safety and task performance over existing diffusion-based planning baselines. |
|
website |
code |
description
Affordable housing is already scarce: nationwide, only 35 affordable and available rental homes exist for every 100 extremely low-income renter households, and just one in four eligible renter households receives federal rental assistance. Even when opportunities exist, families can miss them because applications are buried under confusing requirements and paperwork. RealDoor clears that paper wall. Applicants upload what they have, confirm information extracted with traceable evidence, get cited answers about program rules, catch missing or expired documents early, and export a review-ready packet. AI reads, explains, and prepares; applicants stay in control, and qualified people make eligibility decisions. |
|
code |
description
jaxtaylor is a native JAX high-order adaptive Taylor ODE integrator that reimplements the algorithm of heyoka (bit-exact step size) as a first-class JAX primitive, fully composable with |
|
code |
description
A bottom-up Python implementation of the Finite Element Method: mesh handling, assembly of the discretized system, and solution are all built from scratch rather than relying on an off-the-shelf FEM library. The main focus is on convection-dominated flows, where the standard Galerkin solution develops spurious oscillations, and on SUPG stabilization (streamline diffusion), which adds artificial diffusion along streamlines to recover accurate, oscillation-free solutions. We apply the method to simulate river pollution, tracking the concentration of a pollutant transported by the river flow. Part of the discretization procedure builds on the EPFL Numerical Approximation of PDEs course. |
|