Tieru Wu

Shrinking the entropic regularization parameter in Schrödinger Bridges unlocks a 3x speedup in visual navigation without sacrificing accuracy, finally making diffusion-based policies viable for real-time robotics.

Flow-matching VLAs can ditch 80% of their inference time *without* sacrificing performance, thanks to a clever self-distillation trick that compresses multi-step denoising into a single shot.