Motivation
We cannot directly deploy embodied AI agents in human environments, as they may exhibit unforeseen and erratic behavior under out-of-distribution conditions, requiring formal safety verification. Since human environments are designed for humans, robots operating within them often require human-like capabilities, such as those of humanoids. Verifying the safety of human-like robots is challenging due to their high-dimensional dynamics, potential instability, and real-time constraints. The safety verification of autonomous robots in human environments therefore needs to be tightly integrated with learning-based models for perception, planning, and control. The key challenge is to develop these learning-based verification techniques in a certifiable manner to allow real-world deployment around humans.
Prior Work
My core approach to address this challenge is integrating set-based reachability analysis with deep learning approaches. First, I developed a safety shield for human-robot collaboration that can be applied to any autonomous agent to formally verify human safety in real-time at 1000 Hz. This system guarantees that the robot is either stopped before contact or that the contact forces at impact are below pain thresholds, depending on the chosen mode. I deployed this system in international collaborations with Stanford University and an EU project to enable autonomous robots to perform their desired tasks with minimal safety intervention. My second key achievement in safety for AI agents is the development of the theory for provably safe reinforcement learning. My research on safe RL demonstrates that we can formally guarantee the safety of agents while improving their performance by removing unsafe — and thereby irrelevant — actions.
Research Vision
Building on these achievements, I develop certifiable safety techniques that integrate learning-based methods with abstract system models. To enable certifiable verification based on learning-based perception inputs, I plan to incorporate estimations of aleatoric and epistemic uncertainty into reachability analysis-based verification. For inherently unstable robots such as humanoids, I design RL techniques that learn failsafe trajectories capable of returning the robot to a guaranteed safe state. Finally, I aim to investigate methods for learning the safe action space of RL agents in high-dimensional action spaces to achieve verification of systems for which abstract models tend to be overly conservative. I build on this groundwork to develop verification methods for nonlinear, high-dimensional robots that tightly collaborate with human partners.
Path to Deployment
By developing these methods, I aim to enable the commercial deployment of autonomous robots. I plan to collaborate with small and medium enterprises to introduce humanoid robots into their work environments, initially focusing on mobile manipulators and later advancing toward legged robots. In this context, I am already consulting with the German startup scaledrive.ai to achieve certification for deploying our safety shield on their humanoid robot. In the long term, robot design will increasingly adapt to human needs, emphasizing compliant and soft robots for safe, close interaction. I continue working with industry partners and certification agencies to ensure the rapid and reliable integration of advanced robotics technologies into the European industrial landscape.