Projects & results

General Motors · Cruise Team · 2026

Vehicle fleet motion reconstruction using ML

Scalable scene reconstruction for recorded-driving simulations. A large-scale data pipeline processed terabytes (TBs) of raw tracking data across 38,000+ highway scenes. A distributed training pipeline used multiple H100 graphics processing units (GPUs), PyTorch DistributedDataParallel (DDP), the NVIDIA Collective Communications Library (NCCL), and mixed precision.

Error compared with the team’s prior baseline on approximately 58,000 held-out vehicle trajectories.

Machine learning · Bayesian control

Optimal control through Bayesian inference

Optimal Inferential Control (OIC) reformulates optimal control as Bayesian inference. GPU-compatible ensemble methods solve the resulting inference problem for ML-based vehicle dynamics models, high-dimensional fluids, and thermofluid systems.

Tensor-variate ensemble Kalman smoothing: up to 18.6× faster computation and approximately 11× lower memory use than a vectorized GPU baseline.

Continuous-time control

Continuous-time control of neural ODEs

Continuous-time model predictive inferential control applied to neural ordinary differential equation (Neural ODE) models, including a soft-robot manipulator. Tested configurations ran up to 200× faster than the Interior Point Optimizer (IPOPT) on GPU in tested benchmarks.

Mitsubishi Electric Research Laboratories (MERL) · Energy systems · 2024

Physics-constrained modeling of HVAC systems

A physics-constrained model combining Neural ODEs and gated recurrent units (GRUs) for vapor-compression heating, ventilation, and air conditioning (HVAC) systems. With limited real-world data, it improved forecasting accuracy by 14% and achieved a 5.7× runtime speedup versus the prior modeling approach.

Autonomous vehicles · Planning

ML-based vehicle dynamics modeling and motion planning

ML-based vehicle dynamics models trained on real driving data support Bayesian planning and trajectory optimization. Student’s-t distributions, sequential Monte Carlo, and ensemble Kalman smoothing address non-Gaussian uncertainty and outliers.

IEEE: Institute of Electrical and Electronics Engineers.

Sharif University · Robotics

Marine waste-collection robot

A mobile platform with a selective compliance assembly robot arm (SCARA) combined rapidly exploring random trees (RRT), Dijkstra, and potential-field planning with proportional–integral–derivative (PID) control. The B.S. project included mechanical design, SolidWorks computer-aided design (CAD), and MATLAB/Simulink simulation.