We can't solve AI without talking about safety.
We can't solve safety without talking about people.
I am a Ph.D. Candidate in the Electrical Engineering and Computer Sciences Department at UC Berkeley, in the area of Control, Intelligent Systems and Robotics. My advisor is Prof. Shankar Sastry and I also collaborate with Profs. Anca Dragan, Claire Tomlin, and Tom Griffiths.
I am interested in the challenge of introducing robotics into the public space, allowing autonomous systems to safely and efficiently interact with people. To do this, I seek to understand the interactions between humans and other agents in their environment, bringing together techniques from control and decision theory, machine learning, and cognitive science to design human-centered systems that can leverage synergies and guarantee safety. As part of this effort, I am actively involved in UC Berkeley's Center for Augmented Cognition and Center for Human-Compatible AI.
Bio in a nutshell
Before coming to Berkeley I got my B.S./M.S. Engineering degree at the Universidad Politécnica de Madrid in Spain and an additional Masters at Cranfield University in the UK. I also spent one year working in UAV system design at Aerialtronics. I was then awarded the La Caixa Foundation fellowship to pursue a graduate degree in the United States. At the midpoint of my PhD, I spent 6 months doing R&D work at Apple.
Here are the main research projects I am currently working on.
Click on the images to learn more.
Safe Learning for Robotics
Safety in Multi-Agent Systems
Value-Aligned Artificial Intelligence
The fast advances in learning-based control techniques such as deep reinforcement learning opening exciting opportunities for robots to teach themselves to become better and smarter by interacting with their environments. However, unlike simulated agents, robots often cannot afford the luxury of extensive trial and error, since certain "errors" (say, crashing into a wall) can be unacceptably costly. In other words, robots are safety-critical systems.
In our work, we have looked into principled ways to augment robotic learning with a safety-preserving controller that builds an increasingly reliable predictive model of the robot and its environment and uses it to determine what actions may result in unacceptable outcomes. This controller can be wrapped around any learning algorithm, giving it as much freedom to explore as possible, but overriding it when safety is on the line or uncertainty becomes large due to unexpected environment behavior.
A central challenge in many fields of decision theory (including control theory and artificial intelligence) is analyzing and designing complex large-scale systems with many interconnected elements or agents; applications range from energy distribution networks and air traffic management to systems biology and cancer research. The combinatorial explosion in the complexity of these systems makes their study intractable in many cases; in others, however, structure can be either found or imposed to allow properties and guarantees to be computed.
With the concrete application of unmanned aircraft traffic management in mind, we have introduced a number of theoretical tools and algorithmic approaches to compute tractable strategies for large-scale cooperative multi-agent systems. In particular, autonomous aerial vehicles can travel along naturally-emerging air highways forming platoons, as well as determining near-optimal collision-free trajectories to their destination by sequentially reserving space-time regions. Part of this work is being done in collaboration with NASA's UTM project.
The role of robotics has already begun its expansion from industrial production into the service sector. As robots step out of factory floors and into our homes, workplaces, cities, and roads, performance and safety will no longer be one-sided problems, but will heavily depend on how these machines interact with people. Human-automation systems will become increasingly ubiquitous and complex over the next few decades, and understanding them will be one of the keys to technological and social progress in the 21st Century.
In recent years, cognitive science has advanced our understanding of human inference and decision making through the introduction of mathematical models with remarkable predictive power. We have been theoretically and empirically exploring the integration of these models into the decision processes of robots with some promising initial results.
As AI systems continue to become increasingly capable and complex, predicting their behavior is turning into a major challenge for users and designers alike. Given a user-specified goal, current AI agents can often learn surprising ways to achieve it: sometimes these can be genuinely ingenious solutions that the user had not thought of; in other cases, they can have serious undesired consequences.
In future years, it is expected that AI will come to have an extremely high impact in human society, transforming the way we work, commute, and interact with each other. Both individuals and institutions will rely more and more on AI systems to make decisions for them. To ensure that the impact of AI will be beneficial, we need to find principled approaches to align its decision processes with what people really want; we call this the value alignment problem.