Zi-Yao Zeng

A single 500-hour visual-tactile dataset boosts zero-shot contact prediction from 13% to 38% IoU and quadruples robot manipulation success rates, proving scale and consistency are paramount for embodied learning.

Conditioning video world models on continuous tactile dynamics cuts hand-motion underestimation from 23% to 9%—and surprisingly boosts visual prediction accuracy even when touch sensors are removed at inference.

The Decoupled Embodiment Model (DEM), which pairs a fine-tuned DINOv3 encoder and a frozen NeoBERT encoder with a MeanFlow head that generates each action chunk in a single forward pass, achieves observed success comparable to state-of-the-art VLM-backbone policies under the evaluation protocol.

Task-conditioned foveated perception can drastically enhance the robustness and efficiency of robotic foundation models by aligning policy learning with relevant visual evidence.

Generating missing multi-view data from diverse driving videos boosts closed-loop driving robustness in edge cases by over 30%.

UniTac achieves state-of-the-art tactile understanding and generation by seamlessly integrating sensor and object attributes, revolutionizing how we interpret touch across modalities.

Current AI agents excel in structured tasks but falter at generating novel insights and tackling open-ended scientific challenges.