Humanoid bipedal legs with a natural walking gait.
Two years turning research into robots that walk.
Reinforcement-learning walking policies trained in Isaac Lab and deployed zero-shot on hardware, through a real-time ROS 2 control loop and 2000 Hz CAN.
CPG gait generation, an agentic high-brain / low-brain control architecture, and visual-inertial state estimation.
Fanuc and Dobot automation on the production line.
Backend APIs in TypeScript and Postgres for a non-profit learning platform.
Machines I have built end to end, from a sketch to a field test.
A rocker-bogie rover that vacuums leaves off grass, soil and pavement, and knows where it is by dead reckoning — wheel encoders for distance, a Kalman-fused gyro and magnetometer for heading, and a fine-tuned SAM that reads the ground ahead.
A pen plotter that photographs you and draws your portrait. Face detection and CLAHE clean the frame, Canny reduces it to one-pixel lines, and a circular-sampling vectoriser turns that bitmap into ordered G-code strokes.
A printed hand that copies yours. MediaPipe turns a webcam frame into twenty-one landmarks, a geometric rule reduces those to five bits, and an Arduino drives five servos to match — thirty-two poses over a framed serial link.
Three stepper joints and a gripper running pick-and-place from a phone, on a 3D-printer shield. Open loop: it homes against its own limit switches and then counts steps, with hand-authored joint waypoints instead of a planner.
Bare-metal on an STM32F411. Eight reflectance sensors calibrated per channel and read by free-running DMA, collapsed to one continuous position, steered by a 200 Hz PID — and a one-frame memory that turns a lost line into a recovery instead of a crash.
Flying a DJI Tello by hand. The gesture detector from the robotic hand project, lifted out as a module: the same five bits become an address into eleven drone commands — including one that makes the drone photograph the room, and one that films its own 360° turn.
Handing the meta learning rate of a support/query learner to KerasTuner's Bayesian optimisation, then reading fifteen trials properly — the winners crowd the top of the range, which says the search was bounded too low rather than that it found an optimum.
A CubeSat constellation for watching polar ice: a Walker Delta of twenty 3U observers at 97.8°, a sun-synchronous relay that decides what is worth downlinking, and 873 days of real CryoSat-2 retrievals run through the pipeline the mission would feed.