Jiacheng Wang

Unlabeled robot deployment failures don't need expensive corrective demonstrations to be useful: steering linearly separable latent states away from failure modes fixes policy execution with just 0.5% added parameters.

Real-world robot failures can be autonomously converted into interactive digital twins for targeted RL, boosting out-of-distribution manipulation by 18.75% while completely bypassing manual teleoperated recovery.

Calibrated real-synthetic training can outperform traditional methods in whole-heart segmentation, revealing the power of data-centric approaches in medical imaging.

Coordinating cache compression with a process reward can slash token generation by up to 65% without sacrificing accuracy.