Ching-Hsiang (Tom) Hsu

Robotics AI Systems · Runtime, Replay, Validation, and Deployment

I build the system foundations that make robotics AI reproducible, testable, and safe to deploy. I began in autonomous-driving motion planning, then took on safety fallback, simulation validation, and real-world deployment as those algorithms became production systems.

Today at Woven by Toyota, I work closer to the vehicle platform: deterministic runtime capture and replay, heterogeneous edge-AI pipelines, performance engineering, cross-module integration, and a modular SDK. That progression has shaped how I approach robotics architecture - connecting data, runtime behavior, validation, and safety so complex AI systems can be understood and trusted in the real world.

Overview

Systems Foundation

C++, Python, ROS2, Protobuf, Docker, Git
Low-latency systems, concurrency, profiling, and heterogeneous CPU/GPU computing

Replay and Validation

Deterministic capture and replay, reproducible debugging, simulation-in-the-loop validation, performance benchmarking, and observability

Production Robotics

Motion planning, minimum-risk fallback, edge-AI pipelines, vehicle interfaces, and safety-critical system integration

Architecture Lens

Designing the interfaces between AI models, data engines, runtime platforms, validation systems, and real-time products

Selected Impact

Planning in Complex Cities

Led a production planning framework combining machine-learning and rule-based decisions, improving decision accuracy by 60% in complex urban environments.

Safety Under Degradation

Designed a minimum-risk maneuver framework that achieved a 100% safe-stop rate across more than 10,000 simulated failure scenarios.

Reproducible System Behavior

Built deterministic runtime capture and replay for safety-critical debugging and validation, turning intermittent system behavior into repeatable engineering evidence.

AI at the Edge

Benchmarked and optimized end-to-end imaging pipelines on NVIDIA Orin and Thor across latency, throughput, and memory constraints.