Get rolling with machine learning on AWS DeepRacer
summary
Hands-on demo of AWS DeepRacer Evo shows a six‑layer neural net trained in a 3D simulator (AWS RoboMaker) with models built in Amazon SageMaker, artifacts stored in S3 and video via Kinesis, logs in CloudWatch. Evo adds stereo cameras and LIDAR to let models detect objects. Training captures ~15 images/sec, typically converges after ~20 episodes, and the reward function is authored as a Python Lambda that defines the fixed action space; garage lets you customize sensors and action sets.
AI-written summary · about
chapters (22)
- 0:00Intro
- 0:35About Me
- 3:18Building your team's skills
- 5:23Get hands-on experience with reinforcement
- 6:23Reinforcement learning in the broad Al context
- 7:18Reinforcement learning in the real world
- 7:54Reinforcement learning terms
- 9:26The reward function in a grid race
- 10:18Rewards that incentivize center-line driving
- 10:56Iteration and convergence
- 11:42Exploration vs. exploitation
- 13:39AWS DeepRacer neural network architecture
- 15:35AWS DeepRacer simulator architecture
- 16:27Programming your own reward function
- 17:47Track Components
- 18:51Coordinate system and track waypoints
- 20:09Customize your agent's sensors in the Garage
- 20:39Action space
- 21:04Announcing new 2021 League features
- 21:59Announcing new 2021 Community Leagues
- 22:35Additional resources
- 23:01AWS DeepRacer console
related talks
24mGet Rolling on Machine Learning
27mGet rolling with machine learning on AWS DeepRacer
29mGet rolling with ML on AWS DeepRacer
37mAWS DeepRacer: Train, Evaluate, and Tune
37mAWS DeepRacer:train, Evaluate, and Tune Your Reinforcement Learning Model
37mAccelerate your team’s machine learning journey with DeepRacer
42mFrom Zero to Machine Learning in 45 Minutes
29mKickstart Your Machine Learning Career with AWS DeepRacer