Title:  Efficient Trajectory and Policy Optimization using Dynamics Models

Xinyan Yan

Robotics PhD Candidate

School of Interactive Computing

Georgia Institute of Technology

Date: Monday, June 29, 2020

Time: 2:00 PM - 4:00 PM (ET)

Location (BlueJeans): https://bluejeans.com/459282168

**Note: this defense is remote-only due to the institute's guidelines on COVID-19**

Committee:

Dr. Byron Boots (Advisor), Paul G. Allen School of Computer Science and Engineering, University of Washington

Dr. Sonia Chernova, School of Interactive Computing, Georgia Institute of Technology

Dr. Frank Dellaert, School of Interactive Computing, Georgia Institute of Technology

Dr. Nathan Ratliff, NVIDIA's Robotics Lab, NVIDIA Corporation

Dr. Vikas Sindhwani, Google Brain, Google LLC

Abstract:

Data-driven approaches hold the promise of creating the next wave of robots that can perform diverse tasks and adapt to unstructured environments. However, gathering data of physical systems is often a labor-intensive, time-consuming, and even dangerous process. This issue of data scarcity motivates us to design algorithms that can benefit from prior knowledge that is represented in general forms.

One notable form of prior information is dynamics models. Dynamics models are compact and general: they summarize our knowledge of the robot in the mechanical design and prior interactions with the robot through system identification. Unfortunately, often, the prior information in dynamics models is inaccessible due to limited onboard computing resources and the unavoidable errors in dynamics models are exploited leading to performance bias.

To address this gap, we focus on a central problem in robotics: trajectory and policy optimization; we propose fast inference and prediction algorithms in sparse Gaussian process models, and model-based online policy optimization algorithms that are provably unbiased and sample-efficient. Our research increases the practicality of large-scale kernel machines for real-time trajectory optimization and accelerates policy optimization using dynamics models while avoiding performance bias due to model errors. We evaluate our approaches on a series of robot estimation, planning, and control tasks that involve both simulated data and real robotic systems.