Abstract: Building
robots that can walk or fly with the dexterity of an animal is hard.
Our best models of these robots are complicated because they are
nonlinear and underactuated, and because the robots are fundamentally
subjected to large sources of uncertainty relative to their control
authority – for instance, large disturbances like a gust of wind and/or
model uncertainty because they are locomoting on unknown terrain.
While the field of robotics has seen some success from “intuitive”
controllers based on simple models, I believe that a more algorithmic
approach is required to scale much beyond the complexity and
performance of today’s robots. In
this talk, I will demonstrate that combinations of randomized motion
planning and trajectory optimization can produce open-loop trajectories
for models of the required complexity (nonlinear, hybrid,
underactuated). But when one goes to implement these on the real
systems, they inevitably fail, in part because trajectory motion
planning by itself doesn’t provide any language for talking about
robustness. Therefore, I’ll describe tools from robustness analysis
(both linear and nonlinear), feedback planning, and planning in belief
space that can help close this gap, and produce solutions that work on
real systems. These ingredients can be added on top of most existing
motion planners, and can make a substantial difference when working
with high-performance dynamic robots.
- This event has passed.