Blar Huang Resume

De Anza CollegeCupertino, CA

Computer Engineering, associate degreeSep 2026 – Jun 2028 (expected)

Honors student; Honors project (in progress): uses calculus to show when a drone safety filter (Control Barrier Function) breaks under PX4's velocity-tracking lag (α·τ > 1/4), to be tested in PX4 SITL
ICPC: De Anza College Representative
Competitive Programming Club (CPC): active member
ISMP, De Anza College: Officer Intern
Software Engineering Officer

National Hsinchu Senior High SchoolHsinchu, Taiwan

High school diploma; highly selective public high school (top 5% nationally)Sep 2023 – Jun 2026

Overall GPA: A
Distinguished Student Representative
Ranked 1st of 635 students in the Information Technology and Advanced Programming Curriculum
Peer Counselor and Student Mentor: guided peers on academics, activities and school life
Speaker and Event Host: delivered the annual technology talk to high school peers

Research Experience

NYCU Intelligent Control and Applications LaboratoryHsinchu, Taiwan

Research Intern, advised by Prof. Ching-Hung LeeSep 2024 – Jun 2026

Researched UAV autonomy (motion planning, control and real-time obstacle avoidance) with graduate student mentors
Implemented and compared reactive, sampling-based and optimization-based planners on one PX4 SITL + Gazebo + ROS 2 test stack: Dynamic Window Approach (DWA), Artificial Potential Field (APF), Vector Field Histogram (VFH), nonlinear Model Predictive Control (MPC) and RRT*
Extended DWA into a holonomic variant for multirotors (HOLO-DWA) and identified PX4's velocity-tracking lag from flight logs to tune planners against it
Developed D-Guide (below), a modified-DWA drone navigation system on a ROS 2 physical drone platform, integrating visual perception, waypoint planning and flight control
Designed UAV parts in SolidWorks and integrated flight hardware and sensors

Projects

D-Guide: Voice-Commanded Guide DroneNYCU ICAL

ROS 2 Humble, Python, Pixhawk 6C, Raspberry Pi 5, 2D LiDAR, Google Maps API, Claude API, YOLO, Docker2026

Drone that flies ahead of a walking user and leads them to a spoken destination, aimed at children, older adults and visitors who cannot easily read a phone map
Turned an address into a walking route (Google Geocoding + Directions) and then into GPS waypoints, served as a ROS 2 service
Flew each waypoint with LiDAR-based obstacle avoidance (HOLO-DWA, below) re-planning at ~10 Hz
Voice pipeline: Porcupine wake word → Google Cloud speech-to-text → Claude API parses the request into a structured command, with an offline rule-based fallback
Built on a Holybro X500 V2 frame with 3D-printed sensor mounts; YOLO person tracking to match the user's walking pace

HOLO-DWA: Holonomic Dynamic Window Approach for MultirotorsNYCU ICAL

Python, NumPy, ROS 2, PX4 SITL, Gazebo, Micro XRCE-DDS2026

Reworked the classic DWA to search planar velocity (vx, vy) instead of (v, ω), with yaw decoupled, so the multirotor can strafe through gaps while keeping its forward LiDAR arc on the goal
Tuning study on a 2D-LiDAR course (1.5 m wall gap, cylinder slalom, 2 m gate): baseline reached 1 of 5 runs with 44+ collision episodes; tuned planner reached 15 of 15 across three batches on fresh simulator stacks, 0 collisions, 0.66 m minimum clearance, 18 s average; 1.7 ms per 20 Hz tick (3% of the budget)
Iterated fly → read per-tick logs → find the mechanism → fix: clearance measured along the predicted path rewarded slow candidates and made the drone creep into obstacles, so replaced it with a speed-independent 1.5 m directional probe; a fixed goal bonus made the drone orbit the goal, so replaced it with a braking-curve target speed v = √(2ad)
Traced unexplained collisions to a mirrored LiDAR point cloud (Gazebo's left-positive scan vs. PX4's right-positive frame) that only the asymmetric obstacles exposed; proved it with a frame-check script (0.08 m vs. 0.32 m surface error); also found a ros_gz bridge version mismatch that silently dropped every message

APF: Lag-Aware Artificial Potential Field for a PX4 QuadrotorNYCU ICAL

Python, NumPy, ROS 2, PX4 SITL, Gazebo, OAK-D Lite depth camera2026

3D potential-field planner at 20 Hz on known obstacles or a depth camera point cloud; repulsion from the nearest point in 12×3 direction sectors so a dense cloud cannot out-push the goal, plus a sideways swirl that turns a head-on obstacle into a go-around
Fit flight logs and found PX4 tracks velocity commands as a 0.7 s first-order lag after 0.25 s of dead time, about 2× what the tuning assumed; with that lag modeled, the old tuning clipped pillars in 17 of 24 runs
Moved safety out of the field into a lag-aware brake (v·T + v²/2a = room), which let the field soften (k_rep 1.8 → 0.6) and open 1.3 m gaps; escaped the new local minima with virtual pillars (Park and Lee)
On 55 held-out worlds: 12/55 → 55/55 reached cleanly, worst clearance −0.39 m → +0.41 m; in Gazebo, reached all 7 test worlds vs. 3 of 7 before; a 10 s obstacle memory raised depth-camera runs from 20/55 to 55/55

Nonlinear MPC Obstacle Avoidance for a PX4 QuadrotorNYCU ICAL

Python, CasADi, IPOPT, ROS 2, PX4 SITL, Gazebo, 2D LiDAR2026

Re-solves a nonlinear program every 20 Hz tick: double-integrator model, 20-step × 0.1 s horizon, multiple shooting; p95 solve time 15 ms of the 50 ms budget
Made speed, altitude and obstacle-distance limits soft (slack) constraints so the problem stays feasible after an EKF jump or near-collision, exactly when avoidance matters most; replaced an exact max-based obstacle distance that broke IPOPT's step-size computation with a log-sum-exp smoothing
One solver handles both known 3D obstacles (flies over short ones) and 2D LiDAR columns of unknown height; warm start from the shifted previous solution, brake-to-hover fallback on solver failure, and left/right arc seeds to escape the "brake and stop" local minimum; 18 unit tests

Other Planners: VFH, RRT*, WARP (NYCU ICAL, 2026): VFH+ with VFH* look-ahead reached 56 of 56 unseen test worlds on 2D LiDAR (min clearance 0.8 m); RRT* with LiDAR-map re-planning; WARP, a Riemannian-metric planner with 24 unit tests.

px4-flight-debugger (Personal, 2026): MCP server that lets Claude diagnose PX4 flight logs; traced 2 of 16 failed simulation runs to a compass loss and PX4 failsafe, not the planner.

Orkio: Agent Orchestration Framework for TeamsCal Hacks 13.0 · SpacetimeDB Award

TypeScript, Node.js, Claude Code hooks, MCP, Factory Droid SDK, SpacetimeDB, Git, Vitest, React2026

Distributed multi-agent orchestration across developer machines: an event-driven, pub/sub control plane on SpacetimeDB (transactional reducers, real-time subscriptions, role-based access via authenticated views) syncs presence, intents, WIP snapshots, tasks and conflict state with no separate backend
Architected the system skeleton and integrated all four modules (36 commits; 19/19 live end-to-end tests); built the Claude Code hook layer and MCP tool server, and the LLM planner on the Factory Droid SDK, sandboxed with least-privilege permissions (read-only tools on a disposable worktree, 150 s deadline) that decomposes a request into per-owner tasks, reviews and predicted conflicts, committed atomically as one transaction
Non-invasive WIP sync: snapshots uncommitted working trees to per-agent branches via Git plumbing (no commit, stage or checkout on the dev branch); a pre-merge CI gate merges teammates' snapshots in isolated worktrees and runs tsc + Vitest to catch semantic conflicts that merge cleanly
Deterministic verification over model judgment: conflicts resolve only on bilateral agent agreement plus a fresh green check, with human-in-the-loop escalation (5 min / 12 messages) and a continuously integrated full-stack branch pushed only when green; vs. plain GitHub flow: 78% fewer prompts, 37.5% faster, 5.6% fewer tokens

BikeScape: AI-Scored Safe & Scenic Bike RoutingDAHacks 5.0 · 2nd Place

React, TypeScript, Leaflet, FastAPI, Claude API (vision), Google Street View API2026

Team lead of 5: owned the frontend and integration, and cut scope every 2 hours to ship on time
Built the full-screen map with a Modern (Safe route) ↔ Nature (Green route) switch, street detail cards and a Direct ↔ Scenic slider; merged a teammate's park-discovery prototype into the app

Leadership & Volunteer Experience

Computer Science Clubs AllianceTaiwan

Lead Coordinator2024 – 2025

Founded a new Edge AI Department so students could build practical applications with LLMs
Wrote a 50-page C++ textbook for the club curriculum; mentored 5–6 students through project-based learning
Organized community events and cross-school workshops with 100+ attendees

Global Leadership OrganizationTaipei, Taiwan

Product Management Committee Member / Head of StaffFeb 2025 – Aug 2025

Ran team operations and logistics for a leadership training summit; hosted an online info session and ran social media campaigns to grow registrations

Ministry of Foreign Affairs & Ministry of Education, TaiwanCalifornia, USA

Teen Diplomatic EnvoyAug 2025 – Sep 2025

Selected through multiple rounds of nationwide screening to represent Taiwan in California through international exchange and outreach; recognized as an Outstanding Delegate

Waker International Volunteer (2025): taught and supported children in Nepal and Burkina Faso.

Beyond Taiwan Student Mentor (Sep 2026): mentored a student through international university applications.

Honors & Awards

Skills & Interests

Robotics: ROS 2 (Humble), PX4 (offboard control, SITL, ULog), Gazebo, ArduPilot/MAVLink, motion planning (DWA, VFH, APF, MPC, RRT*), 2D LiDAR, SolidWorks, Arduino

Languages: Python, C++, TypeScript

ML / AI: PyTorch, reinforcement learning (PPO, stable-baselines3), YOLO, CasADi/IPOPT, NumPy, Claude API, MCP, prompt engineering, agentic workflows

Tools: Claude Code, Git, GitHub Actions, Docker, Linux / WSL2, React

Certifications: Anthropic Claude Academy: Claude 101, Claude Code 101, Claude Code in Action, Introduction to Model Context Protocol

Spoken languages: Mandarin Chinese (native), English (full professional) | Interests: Robotics, autonomous systems, machine learning