This project uses reinforcement learning to enable a car to learn on-the-drive, unlike pre-trained models that rely on static data.
A human driver completes three laps on the physical racetrack to establish a benchmark time for the AI to compete against.
The AI's driving behavior evolves from random and unstable after 15 laps to smooth and optimized after 150, showing diminishing returns by 1500 laps.
A recap of the demo's results leads into an explanation of reinforcement learning's core ideas like agents, environments, actions, and maximizing rewards.
Q-learning uses a table of states and actions to store learned values, making it easy to inspect and update the agent's knowledge.
The algorithm's behavior is controlled by key parameters like the learning rate (alpha), discount factor, and the exploration factor (epsilon).
The demo integrates PS4 controllers, an Arduino, the Watson IoT platform, a Node.js backend, and a React.js frontend.
A collaboration with Thyssen Krupp applies these reinforcement learning concepts to a full-size vehicle to learn and adapt its driving style.
The speakers answer audience questions about the importance of data quality, how the car stays on the track, and how the algorithm avoids local optima.
msg
Ismaning, Germany
38:12Daniel Graff & Andreas Wittmann
26:55Jan Zawadzki
48:01Jan Zawadzki
19:24Liang Yu
30:57Teresa Conceicao
31:23Sayed Bouzouraa
28:05Toju Duke
52:10Nico Schmidt
Jobs that call for the skills explored in this talk.
Waymo LLC
Mountain View, United States of America
Waymo
Hamburg, Germany
Trinamics
Utrecht, Netherlands
Tesla Motors
Palo Alto, United States of America
Waymo LLC
New York, United States of America
Waymo LLC
Kirkland, United States of America
Waymo LLC
San Francisco, United States of America
