|
Manipulating Uncertainty: Towards Reliable In-the-wild Dexterity with Less Data
Robotic manipulation is increasingly expected to achieve reliable operation in uncertain environments such as kitchens,
warehouses, and homes, yet performance in these settings remains highly sensitive to latent physical uncertainties that are often difficult to know in advance.
This challenge is especially prominent in contact-rich tasks, where uncertainty in friction, mass, geometry, inertia, or compliance can amplify large differences in how objects move, slip, stick, or deform during interaction. Modern robot learning has made
significant progress by using large-scale data to reduce and adapt to uncertainty during training, but such methods often require far more data than necessary when they cannot deliberately target contact interactions that reveal the physical uncertainties
that are most crucial for reliable manipulation performance. In this talk, I argue that the central challenge in manipulation is not merely the presence of physical uncertainty, but rather the lack of principled methods for deciding which uncertainties must
be reduced, represented, or accounted for to achieve reliable dexterity. My work asks whether robots can become more deliberate about uncertainty by learning what is worth knowing, what can be ignored for the task at hand, and what must be accounted for during
control. I will present data-efficient methods for uncertainty-aware manipulation learning and control, showing how robots can actively exploit contact dynamics to achieve strong learning performance from only a small number of informative interactions. Finally,
I will discuss how explicit reasoning over task-relevant physical uncertainty enables robots to solve manipulation tasks reliably without requiring a fully precise model of the world.
Generalizing Ergodic Coverage for Real-World Robotics
As robotic exploration extends into new domains, robots are entrusted with missions under increasingly strict time and energy
constraints, where success depends on prioritizing search according to each region’s importance. However, in settings like search-and-rescue, where success is critical, prioritized exploration must come with a formal guarantee that no region is left unexplored.
Ergodic exploration offers a solution to this problem by guaranteeing full coverage with time spent in proportion to each region’s value, but existing methods are limited to static, well-defined domains. In this talk, I show how to extend these guarantees
to domains of arbitrary geometry, at any scale, over unbounded time-horizons, and in environments that evolve as the robot explores, so that provable, importance-aware coverage becomes practical in the missions where it matters most.
|
