|
Robots have the potential to become extraordinary tools that extend humanity’s reach, automating tasks that range from
exploring the solar system and the vast oceans to manipulating delicate objects in our homes. Across these settings, robots face a common challenge: they must make good decisions when the world is only partially known, data are scarce, and mistakes are costly.
In this talk, I will present recent work from my group on principled numerical optimization methods that enable robots to reason, act, and learn in these partially observable, remote, and uncertain worlds. I will show how these methods expand robotic capabilities
to support long-horizon ocean exploration by balancing information gathering, safety, and resource constraints; improve precision manipulation by accounting for uncertainty in perception, contact, and dynamics; and accelerate multi-modal (visual) policy learning
by extracting more capability from limited experience. Across these domains, the central theme is that optimization provides a powerful language for building robots that are not merely reactive, but purposeful robots that can plan over long timescales, adapt
to changing conditions, and learn efficiently from the data available to them. Together, these methods point toward a future in which robots can go farther, handle more varied tasks, and learn faster in the real world—bringing robust autonomy to environments
where human presence is limited, uncertainty is unavoidable, and intelligent decision-making matters most.
|
