Hi, I'm Brian. I'm a robotics researcher interested in making dynamic robots do real things in the real world. My expertise is primarily in optimization based planning and control, especially vision-aided walking controllers for bipedal robots. I've designed and impelemented several walking controllers and deployed them on the bipedal robot Cassie.
I'm a PhD student in the DAIR Lab at the University of Pennsylvania, where I'm advised by Michael Posa and supported by the NSF GRFP.
During my Ph.D. I worked as a systems analyst intern at Intuitive Surgical. Before coming to Penn, I received my Bachelor’s degree in Mechanical Engineering from Purdue, and worked as a controls engineering intern at John Deere.
@inproceedings{Acosta2023,
title = {Bipedal Walking on Constrained Footholds with MPC Footstep Control},
author = {Acosta, Brian and Posa, Michael},
year = {2023},
booktitle = {IEEE-RAS International Conference on Humanoid Robotics},
doi = {10.1109/Humanoids57100.2023.10375170}
}
It turns out that a straightforward ADMM implementation is surprisingly performant for solving whole body control QPs that arise in optimization based control of legged robots. As a bonus, we can use the full lorentz-cone constraint for the friction cone, instead of the common pyramidal approximation.
Check out the github repo of my implementation here.