My Mission

I work on foundation models for physical AI and am working toward making embodied AI systems deployable at scale with safety guarantees.

Physical AI systems are becoming increasingly capable, but not yet safe enough to deploy reliably in human environments.

My research builds a central missing layer: a policy-agnostic runtime safety shield that enforces provable guarantees from semantic intent to torque-level execution.

Long-term, I aim to help establish certifiable safety infrastructure for physical AI in Europe and translate this into real-world, deployable systems.

Near-Term Research

Recent progress in foundation models suggests that general-purpose robot autonomy may be within reach. Yet real-world deployment still relies on human supervision, similar to safety teleoperation in autonomous driving.

As OpenAI co-founder and former Chief Scientist Ilya Sutskever remarked, we are shifting from the “age of scaling” back into an “age of research”. Scaling improves capability; it does not provide guarantees. What is required are algorithmic advances that introduce structure and formal guarantees.

A central challenge is ensuring that learning-based control systems (VLA-based or RL-based) operate within certifiable safety envelopes. My work on our safety shield for human-robot interaction (IEEE T-RO, demo) demonstrates how runtime verification and failsafe planning can constrain robot behavior in human environments while preserving performance.

The next step is extending this framework toward foundation-model-driven physical AI systems. This requires:

• integrating semantic safety reasoning with foundation-model-based planners;
• constraining learned policies through runtime safety envelopes;
• validating embodied AI systems in realistic industrial and domestic environments.

The objective is to make physical AI systems deployable at scale through formal safety guarantees.

Long-Term Vision

Already today, LLMs lower barriers to building dangerous systems and agentic AI can automate access to harmful resources. As AI capabilities continue to improve, safety risks are increasingly extending from digital systems into the physical world.

While substantial effort has been devoted to digital AI safety, physical AI safety (e.g., for autonomous robots operating in human environments) remains comparatively underdeveloped.

My long-term vision is to help establish certifiable safety standards for physical AI in Europe and translate them into deployable infrastructure for real-world systems (e.g., humanoid robots, autonomous drones, and human-in-the-loop embodied systems such as brain-computer interfaces). For agentic systems, physical safety cannot be separated from semantic safety: robots must not only act within physical constraints, but also align their behavior with human-defined norms and values.

Achieving this requires close collaboration among academia, industry, and regulators to ensure that embodied AI systems are not only capable, but certifiably safe by design.