Building robotics software and perception for autonomous assembly and other manufacturing applications.

• Improved the pose estimation models that enable seamless bin picking, and explored SOTA zero-shot pose-estimation methods to enable the picking of an unseen part.

• Built Bayesian Neural Networks to allow probabilistic predictions that can give a sense of the confidence of the neural network output, thus enabling the robot to pick the part with the most confident pose prediction, and improving bin picking success rate.

• Explored different grasping algorithms.

Teaching Assistant for Visual Learning and Recognition (16-824) in Fall 2022.

Built an end-to-end learning based unified calibration pipeline using a differentiable robot model and minimized calibration errors in the robot cell.

• Founding engineer at Minus Zero, helped build the camera-based perception stack to perform object detection, tracking and vehicle velocity detection.