## Brian Bittner

**Ph.D in Robotics (2020)**

###### Researcher in John Hopkins University

#### Education

###### B.S. Mechanical Engineering, Carnegie Mellon University, 2016

M.S. Robotics, University of Michigan, 2018

Ph.D. Robotics, University of Michigan, 2020.

I graduated from University of Michigan Robotics Institute. I'm interested in reduction-based approached to robot modeling, planning, and control.

In mobility and manipulation research, building physically intelligent systems is critically dependent on the predictive quality of the models you have access to. Those fond of mechanics will assume structure and conservation laws, friction coefficients and dissipation laws to build models from first principles. Data-scientists try to avoid this altogether by fitting a model from system inputs (i.e. motor torques) to observed outputs (i.e. where the robot goes). The issue with the former is gap between physics-based simulators and reality. The issue with the latter is typically the amount of time required to compute a model from data. We try to bridge the positives of both perspectives:

We use the structure of physics and geometry of locomotion to define a broad class of systems (highly damped) that admit a compact regression structure. This means that for any robot of this class, we have the same simplistic model-fitting problem to solve. This has allowed us to build high fidelity models at unanticipated scales of complexity and speed. Here, we get the problem-reduction and generality provided by the insights of geometric mechanics, and complement it with data-driven tools that model reality, not our assumptions about it. An initial article [1] covers the estimation theory and application for kinematic systems, the second article [2] extending it to high friction or over-damped systems. We extended our results from simulation to hardware by hooking up a wheeled-snake robot to our motion capture system [3]. Most recently, we showcased the generality of the approach by using to teach a walking robot made of tree branches how to navigate a room with just fifteen minutes of experimental data [4].

[1] Bittner, Brian, Ross L. Hatton, and Shai Revzen. “Geometrically optimal gaits: a data-driven approach.” Nonlinear Dynamics 94.3 (2018): 1933-1948.

[2] Kvalheim, Matthew D., Brian Bittner, and Shai Revzen. “Gait modeling and optimization for the perturbed Stokes regime.” Nonlinear Dynamics 97.4 (2019): 2249-227

[3] B Bittner, S Revzen. “Geometric gait optimization with a five-link wheeled snake.” American Physics Society March Meeting, Denver, CO, 2020

[4] B Bittner, S Revzen. “A robot made of tree branches can learn to move in fifteen minutes.” Dynamic Walking, Hawley, PA, 2020