Selected work
Most recent first. Each entry leads with the measurement that mattered, and what it was measured against.-
Quadruped Fall Recovery
The policy was fine. The robot wasn’t the robot it trained against — 216 g heavier, at 77.6% of the torque-to-weight it had learned on.
- 1.4 – 2.1 s
- supine to back on its feet, in 10 of 14 logged recoveries — the other four stalled at 5.4–10.5 s
-
SLAM Walking Robot
Two motors, six bars, and no IMU — a walking robot that still maps an indoor lab and knows where it is inside it.
- 0.211 m
- AMCL localization RMSE on the robot, against a spec of ≤ 0.25 m over N = 35 readings
-
ArmY — Grasp-Pose Selection
Grasp detection offers a hundred ways to pick up a jar. Choosing one is the whole job — and detecting the target accurately turns out not to be the same thing as placing the gripper accurately.
- 7 / 10
- grasp poses within ±1 cm — the requirement the system did not meet, alongside 15/15 target detection
-
Compressed-Air Line-Following Car
Five liters of air at two bar is the entire energy budget — no drive motor, a piston engine I designed, and a lane the camera keeps losing.
- 62.3 m
- travelled on one fill, 86% of the 72.67 m the compressed air should theoretically give
-
RoboMaster Engineer — SCARA Arm
The arm never lifts its own shoulder motors, the beam is a carbon octagon chosen for stiffness per gram, and the lift is two lifts — coarse and fast, then fine and precise.
- 80%
- of the inertia a joint accelerates is the motor's own mass, if you mount it on the joint — so neither shoulder motor rides on the arm
