Shanghang Zhang

Current action-conditioned world models fail to reliably follow diverse off-expert actions, risking the effectiveness of policy learning in real-world applications.

Achieving a mean visual order consistency of 0.9872, Robo-Dopamine 2.0 significantly outperforms traditional reward models, demonstrating its superior ability to handle OOD scenarios in robotic manipulation.

Action-conditioned world models may be failing in fundamental ways, with systematic degradation in simulator fidelity revealed by a new diagnostic framework.

$\omega$-0 enables humanoid robots to seamlessly integrate movement and manipulation, outperforming traditional models by predicting coordinated actions directly from sensory inputs.

The data pyramid framework reveals how the interplay of diverse data sources can unlock new capabilities in embodied agents, highlighting critical gaps in current methodologies.

Achieving high perceptual quality in video compression at bitrates below 0.005 bpp could redefine the limits of efficient video transmission.

By decoupling camera and manipulation actions and training them in a coordinated manner, SaPaVe achieves significantly higher success rates in real-world robotic manipulation tasks compared to existing end-to-end vision-language-action models.

MLLMs stumble badly when asked to reason about safety in lab settings, dropping 32% in performance compared to general knowledge, revealing a critical gap for real-world deployment.