Taoyuan Wang | Autonomous Systems & Robotics Engineer

Overview

Software engineer with 8+ years of experience building infrastructure for autonomous systems, robotics, and computer vision. Currently at Wing (Alphabet), working on safety-critical airspace and operational systems for autonomous drone delivery.

Previously led development of 3D point-cloud annotation systems adopted by 20+ autonomous-vehicle and robotics customers.

Impact at a Glance

10×

Operator capacity

16 → 160 drones

99.99%

Platform stability

while operating as sole backend engineer

20+

AV / robotics customers

on 3D annotation platform

56%

Lower labeling cost

through workflow analytics

Selected Projects

Autonomous Aviation Infrastructure

Wing

C++ Distributed Systems Geospatial Safety-Critical Systems

Backend for Wing’s operator-facing aviation platform — the “Google Maps for the sky” that governs when and where autonomous aircraft may fly.

Problem

Drone delivery must enforce spatial, temporal, and regulatory constraints in real time, across multiple operating regions.

My Role

Built the backend for OpenSky — handlers, validators, testing, data migration — coordinating frontend and backend work across teams.

Technical Work

Translated FAA and international regulatory rules into backend enforcement code.
Built handlers and validators for 4D (3D + time) airspace restriction enforcement.
Integrated real-time aircraft tracking, flight-path visualization, and compliance monitoring with Google Maps and internal data sources.
Added an intelligent layer turning live telemetry into operational guidance (emergency landing, temporary restrictions), plus an AI copilot for on-call incident triage.

Impact

16 → 160 drones per operator; lower labor costs; FAA and CASA approvals unlocked in new cities.
99.99% platform stability as sole backend engineer while shipping roadmap features.
Blackbox monitoring with failure injection for emergency-landing commands (sub-second alerts), adopted by other safety-critical teams at Wing.

OpenSky — Our Drone Airspace Platform wing.com/opensky — opens in a new tab ↗

Wing — A Google X Moonshot x.company/projects/wing — opens in a new tab ↗

AI Multi-Drone Coordination

Wing

LLM / VLA Robotics Multi-Agent Systems Simulation

Natural-language control of multiple drones — intent in, executable multi-drone tasks out.

Problem

Operating several drones still required pre-programmed instructions per task; general-purpose missions needed a more flexible interface.

My Role

Led a proof-of-concept agentic coordination system with iterative planning review, human-in-the-loop escalation, and safety validation.

Technical Work

LLM/VLA planning translates natural-language intent into executable drone tasks.
Fuses real-time sensor data with object detection and tracking into a shared environment model.
Integrated 3D visualization and simulation; text descriptions generate custom 3D test environments.
Logs commands, reasoning, actions, and feedback into a reinforcement-learning dataset for continuous improvement.

Impact

General-purpose multi-drone tasks without pre-programmed instructions.
Low-latency control-model interface designed for real-drone deployment.

Safety-Critical Configuration & Authorization

Wing

Authorization RBAC Distributed Configuration Migration

Re-architected a legacy authorization/configuration system into independently managed domains — with zero downtime and full backward compatibility.

Problem

The legacy auth module, built on Google Zanzibar and Baggins, was complex and failure-prone due to synchronization dependencies; it had to be modernized without interrupting safety-critical operations.

My Role

Led the design and implementation of configuration schemas and validation mechanisms.

Technical Work

Decomposed the monolith into three services: Organization, User, and Member License Management.
Centralized permissions into configuration with schema validation that prevents user error.
Role-based permission checks optimized for low-latency paths (e.g., pilot-issued emergency landing).

Impact

Zero-downtime migration; synchronization failures eliminated.
Operational effort cut from hours to minutes; lower authorization latency.

3D Perception & Annotation Platform

Appen

LiDAR Three.js WebGL Spatial Algorithms Computer Vision

Full-stack 3D point-cloud annotation platform for autonomous-vehicle perception teams — led from prototype to production.

Demo: clicking to select a region in a LiDAR point cloud generates a 3D bounding box via DBSCAN clustering — AI-assisted cuboid annotation: click to select → DBSCAN → 3D bounding box

Demo: 4D point-cloud annotation tracking objects across space and time — 4D semantic segmentation: tracking objects across 3D space + time

Problem

State-of-the-art detection models needed extensive labeling, yet annotating large LiDAR scenes was slow — and heavy scenes crashed browsers.

My Role

Led full-stack development (React/Three.js frontend, Flask/Spring backend) and drove ML-assisted annotation from prototype to production.

Technical Work

Click-to-annotate: fast DBSCAN clustering on k-d trees auto-generates bounding boxes.
GPU shaders for point coloring plus octree frustum culling: smooth rendering from 1M to 10M+ points.
Semantic segmentation for point-level classification; real-time AI-assisted interaction.
Annotator-telemetry dashboards giving operations teams visibility into worker bottlenecks.
Sensor fusion experiments: real-time projection of 3D LiDAR points into 2D camera views; multi-sensor (LiDAR, radar, camera) fusion supporting annotation accuracy.

Impact

35% faster annotation; 20+ AV/robotics clients — flagship product born from a hackathon-weekend prototype.
56% lower labeling costs through workflow analytics.

Credentials

Master's and Bachelor's degrees in Computer Science, NTHU
EB-2 National Interest Waiver (NIW) for contributions to autonomous / uncrewed systems and robotics
Google C++ Readability