Hangjun Ye

Ditch the clunky text-based reasoning: LaST-VLA achieves new benchmarks in autonomous driving by thinking in a physically-grounded, latent spatio-temporal space.

Diffusion models can achieve a 10x performance boost in real-world autonomous driving when optimized for trajectory representation, loss space, and augmented with reinforcement learning.

By explicitly disentangling BEV features into semantic subspaces and assigning them to specialized experts, SEF-MAP achieves state-of-the-art HD map prediction, even when sensor data is degraded.

VLMs can now leverage the power of 3D geometric understanding for autonomous driving tasks thanks to a simple plug-and-play module.

Keypoint detectors can now be trained with RL to directly optimize for long-term trackability across image sequences, leading to significant improvements in downstream 3D vision tasks.

Forget quadratic complexity: UFO reconstructs 16-second driving scenes in half a second by unifying feed-forward and optimization-based methods.

Ditch the discrete anchors: MeanFuser achieves state-of-the-art autonomous driving trajectory generation by using a continuous Gaussian Mixture Noise representation and a mean-flow formulation for faster, more robust planning.

By decoupling generation and refinement experts within a masked diffusion VLA model, DriveFine achieves both flexible decoding and self-correction for autonomous driving.

Achieve state-of-the-art performance in vision-language-action tasks with Xiaomi-Robotics-0, a model that executes smoothly in real-time on real robots using a consumer-grade GPU.