Siddhant Haldar

Presents Point Bridge, a pipeline leveraging domain-agnostic 3D point representations and VLM-based extraction to enable zero-shot sim-to-real transfer and improved sim-and-real co-training for manipulation policies.

Introduces Point Policy, which uses human demonstration videos and semantically meaningful 3D key points to learn robot policies without teleoperation data; demonstrates strong real-world task performance and generalization.

Work describing VisuoTactile Local (ViTaL) policies and related research on combining vision and tactile sensing for precise, contact-rich manipulation.

FISH: an imitation learning approach that learns robust visual skills from less than a minute of human demonstrations and uses demonstration-matching rewards to adapt policies across morphologies and camera views.