Three USC faculty members will lead four research projects funded by the Toyota Research Institute (TRI), spanning humanoid robotics, autonomous driving and social barriers to artificial intelligence (AI) adoption in office environments.

The studies are part of TRI’s latest University Research Program (URP 3.0), a five-year initiative supporting 69 research projects from 31 universities across the nation.

Launched in 2026, TRI’s University Research Program aims to foster industry-academia partnerships to address some of the most pressing challenges in automated driving, robotics, human-centered AI and related fields. The initiative reflects TRI’s commitment to advancing research and collaborating with leading research institutions like USC Viterbi School of Engineering.

Bringing together 88 TRI researchers and 104 faculty members, the projects are structured for deep collaboration, with each project co-led by a university researcher and a TRI co-investigator working as peers to ensure that fundamental research and real-world applications evolve together.

The selected USC principal investigators are Yue Wang, Stephen Tu and Mayank Kejriwal, with Wang leading two projects. All three researchers hold appointments at USC Viterbi and the USC Mark and Mary Stevens School of Computing and Artificial Intelligence. Wang is an assistant professor of computer science, Tu is an assistant professor of electrical and computer engineering and computer science, and Kejriwal is a research associate professor at the Daniel J. Epstein Department of Industrial & Systems Engineering and a principal scientist at the USC Information Sciences Institute. Tu holds joint appointments in the Ming Hsieh Department of Electrical and Computer Engineering and the Thomas Lord Department of Computer Science.

Awarded a combined $3 million in funding, the USC-led projects officially launched last month and will continue for three years.

Titled “World Models for Next Generation Autonomous Driving Policy Learning,” one of Wang’s projects focuses on making autonomous vehicles capable of navigating safely anywhere in the world by teaching them to reason through unfamiliar and extreme road scenarios. His second project, “Reasoning in Motion: Dynamic Adaptation and Intelligent Execution of Large Behavior Models,” focuses on developing humanoid robots that can learn and adapt in real time through daily interactions and feedback within home environments.

Tu leads the project “Calibrated Off-Policy Evaluation for Large Behavior Models,” which focuses on creating a smarter digital testing tool that predicts how well a robot will perform new tasks before it is ever switched on, potentially saving significant time and manual labor.

The study “(AI)nthropology in Action: An Ethnographic Study of Human-AI Co-Evolution and Design in Applied Research Administration” was originally proposed by Adam Russell and is now led by Kejriwal. The project examines the human side of technology to understand what prevents office staff from adopting AI and how to design tools that administrators find genuinely helpful and trustworthy.

University Research Program 3.0 Kick-Off Event at TRI’s Headquarters (Photo Credit: Toyota Research Institute)

When self-driving cars get stuck: training autonomous vehicles to handle the unexpected

Have you ever seen self-driving cars like Waymo vehicles get stuck and unable to move in the middle of the street for hours, causing traffic jams?

Whether it’s because a basketball accidentally rolled into the street, lane closures or a person directing traffic on the road, autonomous cars often stop indefinitely or get “stuck” when they encounter a situation their sensors don’t recognize, or even get into accidents when they face scenarios they have not been specifically trained for.

This high-profile failure is exactly what Yue Wang’s team aims to solve in his study.

The project aims to move away from the industry standard of requiring extensive, city-by-city data collection and retraining, and instead the team is developing “world models” that incorporate human-like reasoning into the vehicle’s decision-making.

This addresses the challenge of having autonomous vehicles trained on every single scenario in the world, which is ultimately impossible.

By teaching AI to understand environmental context, such as realizing that a basketball rolling into the street likely means a child is close behind, the research aims to create self-driving systems that can navigate extreme “out-of-distribution” scenarios safely anywhere in the world, without hitting obstacles, coming to a dead stop or encountering scenarios they have not seen.

Imagine a world with self-driving cars that can think and adapt when making decisions on the road, just like the human brain. That would mean safer roads and fewer traffic jams!

Developing humanoid robots that learn as they move

Robots today still lack the ability to learn from real-time feedback and adapt, a bottleneck researchers have faced for years.

Current robotic foundation models are typically trained and then deployed without the ability to adapt afterward, which makes it difficult for robots to handle unpredictable real-world environments.

Similar to the challenge with self-driving cars, scaling up training data to cover every possible scenario is nearly impossible.

This makes training robots to think and adapt on the spot a much-needed key breakthrough, which is the goal of Wang’s team’s project.

Wang aims to revolutionize how humanoid robots function by moving away from the traditional “train and then deploy” model.

His team will develop AI models that can adapt at “test time,” or deployment time, through real-world interactions. To allow humanoids to execute tasks more intelligently, the project focuses on training them with enhanced motion understanding and real-time perception, using improved perception to help the robot understand how to move and interact with the world more naturally, rather than just repeating a preprogrammed sequence.

He also plans to implement real-time learning with feedback from both the environment and human users.

Finally, his team will utilize advanced software and hardware co-design and improved motion understanding to help robots execute tasks more intelligently. Instead of treating the robot’s “brain,” which is the software aspect and “body,” the hardware aspect, as separate entities, this approach integrates them to allow the robot to adapt at “test time,” meaning it can adjust its behavior while it is actually being used in a real-world setting. This is a key part of a critical breakthrough for humanoid robots.

This study specifically targets applications in home settings, and the findings could allow robots to intelligently navigate the unique and changing layouts of complex household environments.

Speeding up how robots are tested and evaluated and improving robot brains

Unlike human brains, robot “brains” are essentially pieces of code or AI models, historically. These codes and models serve as the decision-making center that observes the environment and decides what action to take next.

In robotics, this robot brain is referred to as a “policy.”

Modern policies are data-driven. Like the human brain, they don’t follow a rigid checklist. Instead, they pattern-match and emulate behavior based on large amounts of data to predict what a human would do in a specific situation, such as washing dishes. In the past, robot brains were rule-based—meaning if a robot saw “X,” it would do “Y.”

While humans instinctively know how to move their arms to a faucet or grab a sponge without thinking, a robot requires explicit, complex software to emulate these instinctive human movements.

Even the most advanced “robotic brains” have blind spots and failure modes, like hitting an obstacle or stopping indefinitely.

Today, to ensure these “brains” work and to find failure modes or statistical blind spots, researchers must test these policies by physically running them on robots over and over again to see if they work as intended.

This manual, time-consuming and labor-intensive process to ensure the quality and safety of new robotic policies is a key challenge facing robotics engineers.

Instead of going straight to the physical robot, Tu’s team is developing off-policy evaluation tools. These tools will use existing databases and logs of past robot evaluations to estimate how well a new policy might perform.

This allows researchers to test a large “bank” of potential policies digitally first. They can then identify only a small set of the most promising “candidate policies” to actually test on physical hardware. This ultimately reduces the number of tests needed and creates a prioritized set of policies to evaluate.

This tool could provide a faster, more effective way to confirm that a newly designed controller is a genuine improvement over previous versions before it is ever deployed.

By making evaluation simpler and more effective, this research speeds up the development of robots capable of complex manipulation tasks. In the real world, this could help robots learn to navigate kitchens or bedrooms to assist with daily chores.

As Tu explained, this domain-agnostic technological advancement has broad potential applications, including in household robotics, where it could help to accelerate the development of assistive robots for elderly care.

Why office staff resist AI tools: Barriers to workplace adoption

While AI tools have integrated into our daily lives, many people do not use them for work in office settings. However, there is a lack of understanding regarding the human and social barriers that prevent people, specifically administrative staff,from adopting them.

Kejriwal explained that such resistance to AI use often stems from intimidation by the technology and is driven by a lack of trust due to fear of job loss or workflows that are simply too busy to accommodate learning new tools.

Led by Kejriwal, this team aims to understand why through large collections of data in an ethnographic study, which is an approach using qualitative research methods where investigators immerse themselves in a specific community or cultural group’s natural environment. Based on the study’s findings, Kejriwal aims to help office workers better transition to an “AI-forward” mindset, instead of fear-oriented mindsets, and ultimately embrace AI technologies as useful tools at work.

His team will conduct ethnographic and qualitative social science research through in-depth interviews and focus groups to identify specific barriers to AI uptake. The study will be conducted on a target group here at USC, which consists of a specific group of 10 administrative staff members at ISI to understand their daily workflows and their relationship with AI tools. This will include a mix of 30 to 60 minute sit-down interviews using open-ended questions to gather detailed insights into administrators’ perceptions, fears and challenges regarding AI.

Kejriwal also plans to identify organic interventions, such as specialized training or identifying “champions,” to help staff transition to an “AI-forward” mindset without using a fear-driven approach.

The team also plans to build an AI co-ethnography platform where AI itself acts as a coach and interviewer to help collect and analyze data on institutional workflows.

Kejriwal emphasized that this study’s goal is institutional transformation. He aims to help research organizations, like USC ISI, improve efficiency by integrating AI into their administrative operations in a way that staff find useful and trustworthy.

Published on June 25th, 2026

Last updated on June 25th, 2026