Toyota Research Institute

Shared Decision Making: Partner-Aware Planning When Interactive Human Data is Scarce

Time-critical human-AI teaming – framed as a partner-aware planning problem under uncertainty over how the partner will respond – is a central challenge for assistive driving and for human-robot collaboration more broadly. In this spotlight, I describe how my work at Toyota Research Institute builds AI teammates for time-critical interaction, and explain why existing approaches fall short of acting as true teammates. The case study is Dream2Assist, a model-based RL approach to shared decision-making in urban and racing driving, deployed in simulation and on physical platforms. I show how algorithmic choices in our stack work around the absence of large-scale, real interactive human data. I then lay out four opportunities for reshaping this paradigm and, looking forward, for realizing partner-aware teaming atop robotics foundation policies in factories, homes, and beyond: (i) world models that generalize along human behavioral axes from tiny datasets; (ii) partner-state as a separable, inspectable representation; (iii) long-term co-planning over heterogeneous physical and communicative actions; and (iv) interaction-conditioned uncertainty for safe human-robot collaboration. I close the talk with the modes through which we engage academic groups on these open problems, together with co-developed infrastructure, benchmarks, and the community we hope to grow around them.

MathWorks

AI-Enhanced Control System Development Workflows: An Industry Perspective

AI is increasingly being adopted across the control system development workflow, from system identification and virtual sensor modeling to data-driven controller design. Techniques such as reinforcement learning and data-driven model predictive control offer new possibilities for control design, with RL naturally handling high-dimensional observations such as images, and data-driven MPC enabling constrained optimal control for complex nonlinear systems using learned prediction models. However, several challenges still hinder widespread industrial adoption, including selecting the right control approach for a given application, sim-to-real transfer, hyperparameter sensitivity, and limited formal guarantees for stability and safety. Emerging agentic AI workflows can help address some of these practical challenges by enabling rapid prototyping across control approaches and streamlining the path from design to deployment.

In this talk, we will discuss trends, applications, and challenges we have observed from our interactions with customers at MathWorks. We will also explore how agentic AI, combined with Model-Based Design, can orchestrate end-to-end workflows spanning requirements, modeling, controller design, testing, and deployment.

PlusAI

Scalable Autonomy for Long Haul Trucking

In recent years, autonomous trucking has moved from R&D to real-world commercial deployment. PlusAI is one of the leaders in this domain working with OEMs like the Traton group ( International, Scania and MAN), Hyundai and Iveco. In this talk, we will discuss PlusAI's software architecture that leverages recent advances in deep learning to provide a robust, scalable, and interpretable foundation for safe and generalizable autonomous driving. The architecture comprises three layers -  – Vision language models (VLMs), End-to-End models and Safety Guardrails. We’ll share how VLMs help with meta decisions and handle “long tail” problems, End-to-End models enable scalable deployment across diverse regions and vehicle platforms, and Safety Guardrails provide sanity checks to prevent radical maneuvers.