Andrea Bajcsy, assistant professor at the Robotics Institute in Carnegie Mellon University’s School of Computer Science, has earned the National Science Foundation (NSF) Faculty Early Career Development (CAREER) award. The CAREER Program grants some of NSF’s most prestigious awards to early-career faculty who show exemplary dedication to the mission of their institution and act as academic role models.
Bajcsy’s project, titled “Formalizing Open World Safety for Interactive Robots,” attempts to expand our current understanding of how robots make safe decisions. The five year grant will support research on robot safety “beyond collision-avoidance” in real-world environments such as homes, city streets and hospitals, with a particular emphasis in helping robots act reliably in the full complexity of real-life situations.
“Humans have a nuanced understanding of safety, like stabilizing hot beverages when we are moving quickly with them, asking for help when we are confused, or handling fragile objects delicately,” said Bajcsy. “However, these types of safety behaviors are very difficult to model mathematically for a robot. Our research aims to enable robots with a generalized understanding of safety, just like we humans have.”
Bajcsy and her Ph.D. student Kensuke Nakamura have already made an exciting step towards broadening the scope of robot safety, including introducing a new method called “Latent Safety Filters”, which will be presented at the 2025 Robotics: Science and Systems (RSS) conference. Their method helps robots understand and prevent hard-to-model safety risks directly from visual inputs, like RGB cameras. By performing safe control inside the “imagination” of a generative “world model”, the robot can detect and proactively prevent unsafe outcomes that are very difficult to write down by hand, such as spilling the contents of an opened bag, seen below.
“We were really surprised by how reliable the Latent Safety Filter was even when interacting with novel Skittles bag colors and an unseen change in the background,” said Bajcsy. “While we still have a long way to go to understand the generalization of latent safety, it’s an exciting initial result.”
The NSF CAREER grant places great importance on both research and education. Bajcsy has developed plans to help others gain a more nuanced and practical perspective on robot safety, especially in the era of robot learning. Her education plan involves working with K-12 students on robot programming and troubleshooting, collaborating with government and industry stakeholders to establish broader adoptions of safe robotics practices and supporting the design of a “robotics red-teaming challenge”, where college students design robot safety algorithms and stress-test them. Bajcsy gave a rundown of the red-teaming concept.
“Red-teaming is about deliberately trying to break systems to see where they fail and where they succeed. While this paradigm is frequently used in cybersecurity, I think it will be critical for stress-testing modern robot pipelines, which increasingly rely on complex models like large language or vision models and world models. The red-teaming challenge will give students hands-on experience designing and breaking robot safety algorithms as a way to improve the safety of next-generation robot learning pipelines.”
Through her education outreach and her continuing research, Bajcsy is building a firm foundation for a lifetime of leadership– a crucial element sought after and supported by the National Science Foundation.
“ The NSF CAREER grant is a great confidence boost: it lets me and my lab know that we are heading in the right directions with our research!” said Bajcsy. “It motivates us to keep pushing forward and do great science.”
For More Information: Aaron Aupperlee | 412-268-9068 | aaupperlee@cmu.edu