Supervisor: Prof. Jeannette Bohg

• Designing and implementing a cross-embodiment scheme to zero-shot transfer a policy trained on videos of humans performing a task to a robot. To be submitted in January 2025, aiming for RSS 2025.
• Evaluated Reinforcement Learning methods on robotics tasks that require fast reactive motions in Mujoco. This project is funded by Toyota Research Institute.
• Conducted joint torque feedback analysis on a large-scale robotics dataset - DROID dataset. Presented important rules of haptic data collection in future large-scale distributed robotics dataset at Stanford cross-labs robotics meeting.

• Designed and implemented an end-to-end deep learning-based 3D gaze estimation algorithm. The algorithm is robust to head motions, and it improves the gaze estimation performance by 84.5%.
• Generated more than 100k synthetic images with suitable domain randomization in Blender for gaze estimation training.
• Designed real-world gaze estimation data collection pipeline and conducted data collection. Conducted detailed analysis and visualization of the dataset.
• Implemented a semi-auto labeling tool for pupil localization and segmentation using SAM2.

Supervisor: Prof. Allison Okamura

• Designed and Implemented a force-aware autonomous tissue manipulation model using imitation learning with da-Vinci Research Kit (dVRK). The task completion rate of autonomous tissue retraction increased 50% with haptic sensing.
• To be submitted to RAL around December 2024.
• Presented force-aware autonomous surgery at Stanford Human-Centered Artificial Intelligence Conference 2024.

• Co-led the team in the ICCV'23 Open Fine-Grained Activity Detection Challenge.
• Won third place on the video activity recognition track and second place on the video activity detection track.

Supervisor: Prof. Man-Wai Mak

• Implemented deep speaker embedding for speaker verification with a domain loss to alleviate the languages mismatch problem.
• The performance of the ECAPA-TDNN (pre-trained using the English dataset) on the unlabelled Chinese dataset has improved by 10% with the MMD-based domain loss. Won the Honours Project Technical Excellence Award.

Supervisor: Prof. James Forbes

• Designed a finite-horizon LQR control of UGV for path tracking.
• Robot Operating System was used during implementation. The state of UGV was represented as an element of direct Euclidean isometries, SE(2).

Supervisor: Prof. Yuxiang Sun

• Developed a deep learning-based integration of monocular visual odometry and multi-object tracking.
• Deployed deep optical-flow estimation for localization and 3D object detection models for 3D multi-object tracking.