Estimation and safe control for aerial, mobile, and manipulation robots. I put classical, optimal, and learned methods on the same benchmark and report what actually wins — including the negative results.
 |
 |
 |
 |
| vision-only ship-deck landing | offshore recovery | manipulator pick-and-place | decentralized drone pursuit |
| Project | What it is |
|---|---|
| drone-landing-moving-platform | Vision-only quadrotor landing on moving ground vehicles & seakeeping ship decks — JONSWAP wave-driven decks, ship air-wake, HJ-reachability + CBF safety shields, honest negative results |
| autonomous-go2-inspection | Full ROS 2 Jazzy autonomy stack: a Go2 quadruped autonomously inspecting a facility — Nav2, RTAB-Map, YOLOE gauge detection |
| openarm-control | Bimanual OpenArm v2 platform in MuJoCo: compliant control, cloth folding, articulated objects, ACT policies — unified in OpenArm-Bench |
| swarm-autonomy | Decentralized GPS-denied drone swarm: onboard VIO, bandwidth-limited shared mapping, cooperative pursuit — no central node |
| intercept | Classical vs optimal vs game-theoretic vs learned guidance on one field, ranked with an Elo/Bradley–Terry ladder (simulation-only, educational) |
| manipdyn | UR5e planning & control lab: 8 controllers and 5 planners, fairly compared on auto-tuned, reproducible scenarios |
Real hardware: norma-robot — LLM-supervised 8-DOF robot arm, winning entry of the "AI-Powered Robot Control" track at the Berlin AI × Robotics Hackathon · Unitree Go2 quadruped — hands-on work at the same hackathon (autonomous-go2-inspection is the simulated inspection stack) · Steady-Stride — a physical self-balancing robot (MPU-6050, real-time PID in C++) · eyrc22_Pharma_Bot — Raspberry Pi hospital-delivery robot, e-Yantra 2022–23
