MEng student in Control of Robotic and Autonomous Systems at UC Berkeley (May 2027); capstone at the Agile Robotics and Perception Lab (Prof. Giuseppe Loianno). I work on model-based control and learned dynamics models for robots.
- Fixed-Wing: fixed-wing simulator where SAC policies fly the control surfaces; in wind, the best of 3 seeds completes 99 of 100 test flights, a PID 97.
- turtlebot-nmpc: acados NMPC that predicts lidar-tracked obstacles 3 s ahead; in Webots it avoided a person crossing at 0.5 m/s in 6 of 7 runs, Nav2 MPPI in none of 3.
- WingJEPA: latent world models of a fixed-wing UAV; in calm air, a physics head predicts open loop within tolerance for a median 11.4 s, plain JEPA 1.9 s.
- GeneticWing: flatness-based soaring trajectory planning in simulation, with a flow-matching student that plans in 30 to 50 ms.
- ur-pen-signature: a UR7e grasps a pen with residual PPO and signs with force control, in MuJoCo: 84% grasp success under camera error, against 74% scripted.
- PIXIE: reimplementation of PIXIE (6D pose from one RGB photo and a CAD model), plus a per-part assembly check.
Also: quadrotor-ros2, ur5e-neural-kinematics.