This semester project focuses on enabling an ANYmal with a 6-DoF Arm to open and traverse different types of doors without task specific knowledge. Using student-teacher distillation, we distil the knowledge of task specific expert teachers, trained using PPO, into a single student policy. We build upon this formulation to explore methods to improve task specific feature extraction to improve learning performance.

As lead for systems architecture and integration, I coordinate between subteams and drive critical design decisions to get ETH’s first competition rover ready for the European Rover Challenge

We used Reinforcement Learning to train a control policy for a “camera” drone to track an “actor” drone using vision information. We formulated a custom reward scheme to prioritize tracking, bounding box centering, smooth actions, and safety constraints to achieve reliable camera based tracking.

We implemented a Monocular Vision Odometry (VO) pipeline in Python. Components include landmark and camera pose initialization along with continuous pipeline for keypoint tracking and landmark triangulation using 2D ↔ 3D correspondences. To ensure stable and long-term operation, new keypoint ↔ landmark correspondences were continuously evaluated and added to the main pipeline once deemed to be of high-quality. Evaluated on KITTI, Malaga, and Parking Datasets

Our team designed and fabricated a tendon-actuated hand mounted on a Franka Emika arm for teleoperation, and an autonomous object manipulation task. I was responsible for the base ROS 2 software, hardware interfacing, and developing an Action Chunking Transformer based Imitation Learning policy to perform pick-and-place tasks. I also helped develop a UI with safety checks to prevent hardware faults and accelerate data collection.

The project was recognized with “Most Intuitive Software Design”.