Data-Driven Methods to Build Robust Legged Robots

For robots to ever achieve serious autonomy, they need to be able to mitigate an uncertain environment and variation of their own morphology.

Legged robots, with their complex dynamics, present particular challenges in producing robust locomotion – hybrid events, uncertainty, and high dimension are all confounding factors for direct analysis of models.

On the other hand, direct data-driven methods have proven to be equally challenging – the  high dimension and mechanical complexity of legged robots have proven challenging for hardware-in-the-loop strategies to exploit without significant effort by human operators.

We advocate that we can exploit both perspectives by exploiting qualitative features of class of mathematical models applicable to legged robots, and use that knowledge to strongly inform data-driven methods.

We show that the existence of these simple structures can greatly facilitate robust design of legged robots from a data-driven perspective.

We begin by demonstrating that the factorial complexity of hybrid models can be elegantly resolved with computationally tractable algorithms, and establish that a novel form of distributed control is predicted.

We then continue by demonstrating that a  relaxed version of the famous templates and anchors hypothesis can be used to encode performance objectives in a highly redundant way, allowing robots that have suffered damage to autonomously compensate.

We conclude with a deadbeat stabilization result that is quite general, and can be determined with zero dynamic information.

Chair: Professor Shai Revzen