Research

Currently working on a multimodal grasp prediction system that combines visual data and tactile feedback to forecast the success probability of a robotic grasp, conditioned on the robot’s intended action for future time steps.

By integrating sensory inputs, the system learns robust representations of object properties in unstructured environments.​

Developed perception pipeline for a mobile manipulator designed for aircraft cabin cleaning, leveraging monocular camera input in NVIDIA Isaac Sim.
– Implemented semantic segmentation using SAM and DINO to identify objects for downstream pose detection.
– Integrated GraspSplats for grasp point detection and FoundationPose for 6D pose estimation, enabling precise robotic manipulation. Utilized Mast3r to generate depth maps, enhancing the robot’s environmental understanding for improved motion planning and obstacle avoidance