Research overview

Many systems contain latent variables that make their dynamics partially unidentifiable. More fundamentally, data-driven techniques can confuse correlation with causation, producing spurious distribution shifts in observed statistics that corrupt learned models and safety certificates. We develop safety certificates that remain valid under exactly these conditions.

The foundation is a probabilistic invariance framework that states invariance conditions in probability space. It lets risk measures learned from limited data or observation be converted into safety certificates with long-term assurance despite latent risks. Building on it, we used causal reinforcement learning to construct consistently feasible safety conditions from data even under spurious distribution shifts, the first integration of causal RL into stochastic safe control.

The same framework yields certificates that work with incomplete information, such as occlusions in navigation and the absence of centralized communication in multi-agent systems. It produces myopic, affine control constraints with assured long-term safety and exploits low-dimensional risk representations, analogous to biological reflexes, to operate under limited computation. Because the certificates accept modular risk models, controllers adapt to changing environments and can draw on external risk-quantification tools.

Quantifying long-term risk is difficult for three reasons: risk events are rare in data, computation must scale to high-dimensional and multi-agent systems, and risks differ across heterogeneous systems. Our work addresses each of these.

To learn from limited data, we developed physics-informed learning techniques that combine empirical data with physics information. We showed that four distinct classes of long-term risk probabilities are characterized by partial differential equations, and designed B-spline neural architectures that enforce PDE boundary conditions by construction. Analysis and experiments show that this framework infers long-term risk from short-horizon samples with few risk events, generalizes beyond the sampled states and horizons, and computes stable probability gradients.

For computational efficiency, low-dimensional risk representations scale the approach to problems with thousands of dimensions, and a decomposition framework quantifies risk in sparsely connected multi-agent systems through low-dimensional subproblems. Physics-informed reinforcement learning solves probabilistic reachability problems with fewer visits to risky states and with sparse rewards, for example estimating vehicle drift probability from binary safety labels. For heterogeneous systems, a B-spline neural operator architecture provides provable approximation and generalization guarantees.

Uncertainty quantification is essential for the robust deployment of deep learning in physical systems, yet it usually relies on costly sampling or coarse approximations. We developed a sample-free method that characterizes the input–output distributions of neural networks by propagating mean vectors and covariance matrices through the network. The key enabler is an analytic solution for the covariance of random variables passed through nonlinear activations such as Heaviside, ReLU, and GELU.

This result underpins training techniques for Bayesian neural networks based on deterministic variational inference and, more recently, a dual Bayesian inference scheme that reformulates fine-tuning as two coupled Bayesian updates: one incorporates new data as it arrives, the other performs layer-wise inference through closed-form moment propagation. Propagating analytic moments in a single pass gives ultra-low-latency adaptation from a small number of samples, which is what sequential fine-tuning under shifting distributions requires.

To enhance robustness and adaptability, we developed techniques for attack-resilient estimation and control, stabilization of unknown systems with fundamental sample-complexity limits, predictive control with regret guarantees for non-stationary environments, and fast bandit-based meta-learning across diverse environments.

Safety in the long run also depends on how people respond to machines. Through modeling and human experiments, we showed that a worst-case safe control strategy can trigger adverse human adaptation that increases risk for everyone, and demonstrated how to facilitate more desirable adaptation over many interactions. We also built frameworks that predict and influence human behavior for proactive, safe human–robot collaboration.

These methods have been applied to autonomous driving, robots, smart grids and electric vehicle charging, cloud computing, building HVAC systems, and poverty-alleviation policy design, where our work was the first to bring a control-theoretic perspective to microfinance.

Natural language carries context that sensors and models cannot: instructions, preferences, and warnings about latent risks. We are developing frameworks in which large language models supply this context to safe controllers. Our work translates human instructions into constraints and actions while accounting for context and latent risks, adapts probabilistic safety certificates online from natural-language guidance together with Bayesian estimators, and extends language guidance to distributed control of multi-agent systems.

In the other direction, control-theoretic tools inform the development of language models themselves. As a Project Researcher at the Research and Development Center for Large Language Models at Japan's National Institute of Informatics, I apply control-theoretic techniques to support the development of large language models and language-action models.

We are currently developing next-generation Vision-Language-Action (VLA) models and deploying them onto advanced humanoid robots. On the algorithmic front, we develop ultra-efficient fine-tuning techniques that empower these VLA models to master highly complex, real-world tasks. This deployment phase is conducted in close collaboration with Fujitsu Research as part of their pioneering Physical AI initiative, which includes the Fujitsu–Carnegie Mellon Physical AI Research Center launched in April 2026.

Nervous systems achieve remarkably fast, accurate, and robust control despite severe component limitations in speed and accuracy. Understanding this paradox offers design principles for intelligent control systems. We modeled how neurophysiological constraints in latency and information rate impose fundamental limits on sensorimotor performance in layered control architectures, and built an open-source experimental platform to test the predictions.

The study revealed a core principle: diversity among neurons creates diversity-enabled sweet spots (DeSSs) that deconstrain the limitations of individual components. DeSSs explain both how biological systems achieve fast and accurate control using slow or inaccurate components and why heterogeneity is ubiquitous in nature. Extending the model, we showed how microscopic motor constraints give rise to macroscopic behavioral laws such as Fitts' law, providing a mechanistic explanation for the logarithmic speed–accuracy tradeoffs observed across reaching tasks.

Beyond neuroscience, the same techniques reveal fundamental tradeoffs in biomolecular control and inform the design of networked control systems operating over delayed, quantized, and rate-limited communication channels.

Robots operate with limited data, memory, and computing power in environments that never stop changing, so training them once and expecting safe performance forever is unrealistic. We therefore study sequential fine-tuning, in which a system keeps learning as new information arrives, together with uncertainty awareness, so that a system knows what it does not know and acts cautiously in unfamiliar situations. The long-term vision is a unified framework that integrates language understanding, learning, uncertainty estimation, and control, so that anyone, not just engineers, could configure a robot to perform tasks safely and customize its behavior to personal preferences or workplace norms.