Cyrus Huang

Studies scaling laws for autonomous driving by scaling model parameters, datasets, and compute by orders of magnitude. First autoregressive transformer commercially deployed for autonomous driving, demonstrating strong planning and prediction capabilities in complex real-world scenarios through large-scale training.

A structure-guided diffusion model for traffic simulation that injects causal structures to balance realism and controllability, outperforming state-of-the-art methods across key metrics including collision rate, off-road rate, FDE, and comfort.

A contrastive learning method that uses long-term LiDAR-based correspondence to pretrain coherent instance representations from BEV features, improving data efficiency and performance across perception, prediction, and planning tasks.

Uses vision-language models as teachers to provide rich text and action supervision during training without requiring them at inference, leading to significant gains in end-to-end planning accuracy and collision reduction.

Decouples joint prediction into marginal and conditional predictions for interacting agents, achieving state-of-the-art performance on the WOMD interactive prediction benchmark.

Extends the soft actor-critic framework to estimate execution risk and produce risk-bounded policies by conditioning on continuous risk levels, achieving better plan quality and efficiency in complex scenarios.

Fast non-sampling methods for risk assessment of autonomous vehicle trajectories under probabilistic predictions, handling both Gaussian and non-Gaussian models using numerical solvers, Chebyshev bounds, and SOS programming, with validation on Argoverse and CARLA datasets.

A variational neural network that predicts future driver trajectory distributions with confidence estimates, improving physics-based prediction errors by 25% and identifying uncertain cases with 82% accuracy on real-world urban driving data.