Safe, Robust and Adaptive Model Learning for Agile Robots: Autonomous Racing - Robotics Institute Carnegie Mellon University
Loading Events

MSR Thesis Defense

July

24
Wed
Dvij Kalaria MSR Student Robotics Institute,
Carnegie Mellon University
Wednesday, July 24
9:30 am to 11:00 am
1305 Newell Simon Hall
Safe, Robust and Adaptive Model Learning for Agile Robots: Autonomous Racing
Abstract:
In recent years there has been a rapid development in agile robots capable of operating at their limits in dynamic environments. Autonomous racing and recent developments in it also spurred by competitions such as the Indy Autonomous Challenge, A2RL, and F1Tenth have shown how modern autonomous control algorithms are capable of operating racecars at their handling limits and also possibly capable of beating humans in the near future.
What becomes challenging for such agile robots is their need to adapt rapidly to changing environmental conditions like surface change due to tire degradation, weather change, tire temperature changes in racing; and also, the need to maintain a high control frequency necessary to execute such demanding tasks. The robots also need to stay safe while executing such tasks. Prior works who have approached this problem either assume known parameters with uncertainty; or assume constant parameters during the run; or are compute-intensive and disregard practical delays when deploying on hardware
In this work we show how we can leverage safe learning for agile robots operating at their limits, which safely enables them to learn and exploit their limits while adapting to environmental changes; and at the same time is feasible enough to be deployed on hardware taking care of practical delays. We take autonomous racing as an application example but will formulate an approach which can be applied generally to any agile control system.
In the first half of the talk, I will discuss the formulation of a safe, robust and agile model-based approach for this task. In the second half, I will discuss a proposed robust model-based safety filter for model-free reinforcement learning.
Committee:
Prof. John M. Dolan (advisor)
Prof. Guanya Shi
Simin Liu