Hi, I'm Paul

A team hackathon build on top of the BracketBot platform: an autonomous waiter that docks itself, lifts a tray with both arms, delivers it, and takes voice commands — built end-to-end in 24 hours with Oscar Liang, Oscar Low, and Shiv Kanade.

• Extended the pre-existing BracketBot base with a custom docking station and tray, enabling fully autonomous pickup and delivery with zero teleoperation.
• Built ArUco-tag visual navigation: a calibrated fisheye camera solves marker pose via OpenCV PnP, driving a closed-loop visual-servo controller to the tray, dock, and user.
• Solved smooth, spill-free two-arm lifting using the bbos pub/sub framework for coordinated dual-arm control with preset pose interpolation ("ramping"), backed by a retry-with-verification loop that cross-checks real motor encoder positions against target commands until convergence.
• Integrated on-device speech recognition (faster-whisper over WebSockets) so the robot responds to spoken food requests and drop-off confirmations.

A custom 4-DOF parallel-linkage arm developed simulation-first in NVIDIA Isaac Sim: the design was modeled and validated in a physics simulator before hardware deployment, then paired with an overhead camera that localizes objects in 3D so the arm can pick them up fully autonomously.

• Built a full URDF pipeline to bring the Fusion 360 arm design into NVIDIA Isaac Sim, modeling its physics, closed-loop parallel-linkage kinematics, and joint limits in a high-fidelity simulation.
• Used Isaac Sim as a simulated twin of the arm to validate kinematics and joint limits before touching hardware, catching design issues early and reducing costly physical iterations.
• Built an OpenCV vision pipeline with intrinsic/extrinsic camera calibration to unproject 2D pixels into precise 3D world coordinates.
• Wrote a closed-form analytical inverse kinematics solver in Python, closing the loop from pixel to joint actuation in real time.

A 15-DOF bipedal humanoid, designed and 3D-printed from scratch, that learns to walk and recover its balance entirely in simulation before ever taking a step on hardware.

• Designed, CAD-modeled, and fabricated the full 15-DOF assembly using Fusion 360 and FDM 3D printing.
• Exported the assembly into custom URDF and MuJoCo XML environments with tuned physical and inertial properties.
• Trained reinforcement learning policies (PyTorch, Stable Baselines3) for dynamic walking gaits and balance recovery.
• Deployed the trained policy onto an onboard Raspberry Pi driving the servo actuators.

A fully custom 6-DOF robotic arm built to study kinematic modeling and closed-loop control from first principles.

• Designed, CAD-modeled, and built a complete 6-DOF arm from scratch using Fusion 360 and FDM 3D printing.
• Applied forward and inverse kinematics, with command-based control abstractions in C++ separating motion planning from actuation.
• Tuned PID feedback loops for closed-loop position control, achieving ±1.0 mm end-effector repeatability at full reach.

A PID-based autopilot for an RC plane, tuned experimentally for stable autonomous flight.

• Designed and implemented a PID-based control system for autonomous flight stabilization.
• Tuned control parameters experimentally using real-time sensor feedback.
• Analyzed system stability and responsiveness under changing environmental conditions.