HOME / PROJECTS

From research to reality

Three active research lines, written as notes for a technical reader: the problem, the approach, and how far it is from a production environment.

WIZARD
Embodied AI · Manipulation

WIZARD

Industrial robots need re-teaching for every new task. WIZARD uses one-shot imitation learning, developed with UC Berkeley, to adapt from a single demonstration with zero retraining.

~14×
success rate over the one-shot imitation baseline on our unseen-task evaluation set. Benchmark and protocol are in the paper.
1
video demonstration, then autonomy — no per-task teaching pass.
Stage Active research project, not yet in production. Validated on real hardware in the lab.
LEGO
Robotics · ICLR'26

LEGO

A grasping policy trained entirely on randomly assembled shape primitives, which then transfers to real objects it has never seen — removing the in-domain data collection step.

0-shot
grasping on unseen objects
Play
self-supervised data collection
Stage Published method, validated on a Franka arm in the lab. Not yet deployed in an industrial setting.
T-REX
Robotics · Tactile

T-Rex

Dexterous manipulation degrades when contact is uncertain and vision alone cannot resolve it. T-Rex closes the loop between touch and control. A 36-researcher multi-institutional collaboration — our contribution is documented in the author list, not claimed as sole authorship.

Touch
high-resolution tactile feedback
Reactive
real-time closed-loop control
Stage Active multi-institutional collaboration, currently in the research and validation stage.
Showing all projects