About Me

Auto-Driving | Agentic AI | Human-Computer Interaction

Developer Image

Welcome! I'm Han Wang, a Ph.D. candidate in Berkeley AI Research(BAIR) and Berkeley Deep Drive(BDD) at UC Berkeley. My doctoral research focuses on improving traffic flow and urban mobility with connected and automated vehicles using multi-agent system framework. I have an interdisciplinary background in Computer Science and Transportation Engineering, with a keen interest in vision science and generative AI.

I've worked on exciting projects around automated vehicles like developing speed planner for 100 AVs fieldtest, predicting vehicle trajectories and collision risk evaluations, supported by organizations including General Motors, Nissan, Toyota and Allstate. For personal projects, I enjoy working with AR/VR + GenAI technology and currently addicted to the development of Vision Pro.

Hope you find them as exciting as I do!

Featured Projects

Yes, we sent 100 auto-driving vehicles to the morning peak of an open highway.

Congestion Impacts Reduction via CAV-in-the-loop Lagrangian Energy Smoothing (CIRCLES)
CIRCLES Project Image

As the final field test of CIRCLES project, the MegaVanderTest (MVT) was a large-scale traffic experiment conducted on a 14.5-km stretch of Interstate I-24 in Nashville, Tennessee, deploying a fleet of 100 automated vehicles to improve traffic flow and fuel efficiency. The experiment utilized a hierarchical Speed Planner framework, integrating server-side algorithms with vehicle-side execution to dynamically manage variable speed limits in mixed autonomy traffic environments. The MVT demonstrated significant improvements in traffic smoothness and fuel consumption, validating the practical viability of the Speed Planner in real-world conditions.

Learn More →

That math behind CIRCLES: Can learning-based controller work with PDE/ODE dynamic?

Reinforcement Learning-Based Adaptive Speed Controllers in PDE/ODE simulated Mixed Autonomy Traffic Flow
ECC24 Diagram

This research introduces an adaptive speed controller for Automated Vehicles (AVs) in mixed autonomy traffic, integrating a learning-based Reinforcement Learning (RL) controller with sophisticated mathematical dynamics modeled by Partial Differential Equations (PDE) and Ordinary Differential Equations (ODE). The RL algorithm, operating within an Actor-Critic framework, optimizes AV control policies in real-time by interacting with the PDE-ODE model. Numerical simulations demonstrate significant improvements in traffic flow metrics, underscoring the potential of this integrated approach to mitigate congestion and enhance overall traffic efficiency.

Learn More →

Could GAN generate a real frame of future?

Probabilistic Occupancy Risk Assessment (PORA): Motion Prediction using Pre-Trained Generative Model
PORA Project Image

In this research, the conditional generation capability of pre-trained image generative models is proven to be capable of motion prediction of self-driving vehicles. We introduce a novel Probabilistic Occupancy Risk Assessment (PORA) measure, leveraging a generative model to predict vehicle occupancy and assess collision risk in autonomous driving scenarios. The generative model, implemented as a pre-trained GAN generator, creates occupancy heatmaps from traffic data, enabling dynamic and uncertainty-aware risk evaluation. This approach demonstrates enhanced accuracy in predicting potential collisions, offering a robust framework for improving autonomous vehicle safety.

Learn More →

VR bike simulator rebuilding California streets

Study the Bicycle Level of Service(BLOS) of variant scenarios by plugging riders into it.
BLOS Project Image

In the Caltrans funded Bike Level of Service (BLOS) research, we built a virtual reality bike simulator designed to facilitate urban planning and infrastructure development. The simulator enables users to experience and evaluate different bike lane configurations, road designs, and traffic scenarios, providing valuable insights for optimizing street layouts and enhancing cyclist safety. BLOS offers a user-friendly interface and realistic simulation environment, making it an effective tool for urban planners, transportation engineers, and policymakers.

Learn More →

What's the game, then?

Manuplate the simulation / gaming process via run time behavior generation
GROMIT Cover Image

GROMIT, a system utilizing large language models for runtime behavior generation in games, enables dynamic gameplay interactions by generating and integrating behaviors based on player input without developer intervention. Evaluations show GROMIT's effectiveness in specific scenarios, and interviews with developers reveal interest in its potential applications, alongside concerns about quality and integration into workflows. Future work aims to address these concerns by developing guardrail systems to manage the quality and impact of generated behaviors.

Learn More →

Automating That Boring Part of Your Browser Workflow

Automation of browser-based workflows using vision-based LLM
Web Automation Cover Image

Why do robotics scientists value humanoid robots? Because all the infrastructure and tools of current society are intended for humanoids.

We believe that for workflow automation, vision-based LLMs are the equivalent of humanoid robots in robotics because most current computer software and tools are designed for and optimized around human interactions. Vision-based LLMs can interpret and interact with these systems in ways that mimic human behavior, making them highly adaptable and effective in automating tasks across a wide range of applications.

Learn More →

Your Tesla can see more than you do? Not anymore

Boost your visual perception with external sensors via integration of foundation models (FMs) and augmented reality (AR)
AR Cover Image

This research integrates foundation models with augmented reality to enhance driving assistance systems by providing real-time visual aids and interpreting complex driver commands. The system uses advanced data collection and processing techniques to create a cohesive understanding of the environment, enabling accurate 3D reconstruction and amodal tracking of objects. Key innovations include a structured Scenegraph for processing varied natural language commands.

Learn More →

More About Me?

↓ Discover my full résumé, publications and some adorable cat photos! ↓