Bashar Kabbarah

Bashar Kabbarah

Biography

BIO: Bashar Kabbarah is a rising sophomore studying Electrical Engineering and Computer
Sciences at UC Berkeley, where he is a Regents’ Scholar and a member of the SEED
Scholars Honors Program. His first research project, medical imaging for spinal disease, later
published in the Journal of Emerging Investigators, exposed him to the proclivity of highstakes
machine learning models to fail to generalize beyond lab conditions. As an MIT Media
Lab intern in Professor Paul Liang’s Multisensory Intelligence group, he now addresses this
limitation by leveraging multimodality, researching AI systems that combine vision, language,
touch, and even smell to provide richer sensory grounding that improves model robustness.
His past experience spans Lawrence Berkeley National Laboratory, UC Berkeley’s DNA
Sequencing Facility, and industry work at BFAI Semiconductor Solutions, where he improved
computer vision models for defect detection. Bashar intends to pursue a PhD in multimodal
representation learning for safety-critical domains like medicine and manufacturing.

Piezoresistive Palm Reading: Multimodal Alignment and Prediction from Touch
for Dexterous Manipulation
Bashar Kabbarah1, Kaichen Zhou2, Yuxin Ray Song2 and Paul Pu Liang2
1Electrical Engineering and Computer Sciences, University of California, Berkeley
2Media Lab, Massachusetts Institute of Technology

A wide range of sensory signals help dexterous robots manipulate objects reliably. Firstperson
video captures the appearance of objects and interaction but not contact, force, or
grip, leaving gaps that tactile information can fill. The OpenTouch framework aligns touch,
egocentric video, and hand pose in a shared representation using contrastive learning. We first
improve alignment, then ask whether touch predicts how a hand will move next. Replacing
average pooling with a temporal pose encoder improves tactile-to-pose retrieval 2.7x, from
16.8 to 45.5 mAP. To test prediction, we isolate finger articulation from whole-hand motion
and decode its future direction from touch available only up to the present moment. Touch
predicts that direction well above chance, at AUC 0.60 to 0.69, while shuffled-touch controls
remain at AUC 0.50. Adding touch to a pose encoder with the same temporal history still
improves prediction, across horizons from 67 to 533 milliseconds, and helps most when
predicting whether the fingers are about to curl. Touch therefore supplies information that hand
kinematics alone do not carry, revealing not just what a hand is holding but how it is about to
reshape. This suggests robots could use contact to anticipate hand motion rather than react to it.