Yilin Wang

Yilin Wang

Ph.D. Candidate @ Purdue | Cooperative Perception, Physical AI, Autonomous Driving

TeraSim: City-Scale World Model for Safety Testing on AV and Physical AI

  • Role: core developer of TeraSim; Simulation Development Engineer Intern
  • Period: Dec 2025 – Jun 2026 (internship)
  • Affiliation: Inchor (SaferDrive AI), Ann Arbor, MI; joint project with NVIDIA and the City of San Jose
  • Status: ongoing; core developer of TeraSim since Dec 2025
  • Links: NVIDIA press release, Mar 16, 2026 · GTC 2026 news post

Overview

TeraSim is an open-source, city-scale traffic simulation platform for autonomous vehicle (AV) safety testing, enabling naturalistic and adversarial testing of AVs at urban scale. The internship work centered on two parts: NNDE, the naturalistic driving behavior layer of TeraSim, and a joint project with NVIDIA and the City of San Jose on agent-driven Smart City operation. It also included implementing a full-stack Driving Foundation Model into product.

Part 1: Naturalistic Driving Behavior (NNDE)

NNDE (Neural-network Naturalistic Driving Environment) is the behavioral-layer foundation module of TeraSim; it integrates a transformer-based motion model to generate naturalistic driving behavior at city scale.

  • Led the integration of MotionTransformer as the foundation model of NNDE, generating naturalistic driving behavior at city scale alongside the NADE adversarial-scenario module.
  • Refactored the behavior-model pipeline from a dual-state design into a stateless prediction pipeline for large-scale inference.
  • Implemented vehicle state synchronization, history tensor construction, and coordinate transformation for large-scale MotionTransformer inference.

Part 2: City-Scale Demo with NVIDIA and the City of San Jose

  • Led the joint project with NVIDIA and the City of San Jose on agent-driven Smart City operation.
  • For that project, led a city-scale urban simulation demo under the NVIDIA Metropolis Blueprint: an integrated “City Operations Agent” and digital twin that perceive, simulate, and optimize traffic operations.
  • Designed the full simulation pipeline, including city-scale vehicle injection, dynamic routing, and multi-phase demo orchestration.
  • The demo was selected for NVIDIA CEO Jensen Huang’s GTC 2026 keynote (Mar 16, 2026, San Jose, CA).

Tools

Python, SUMO, PyTorch, CUDA, Docker, FastAPI