Replicate AWS DeepRacer architecture to conquer the track with SageMaker
summary
Replicates the AWS DeepRacer stack inside SageMaker notebooks, using RoboMaker for simulation and Kinesis Video Streams for sensor flow, and provides editable network templates (stereo CNN for cameras plus 1-D input for LIDAR) that are ready to run. Experiments show a deep network reaches ~80% track coverage after 400 episodes versus ~60% for a shallow net, and hits 60% coverage about twice as fast, though deeper nets sometimes hurt sim-to-reality transfer due to inference delay.
AI-written summary · about
chapters (12)
- 0:00Intro
- 2:31AWS DeepRacer architecture diagram
- 3:28What does Amazon SageMaker enable beyond the AWS DeepRacer console?
- 5:58Rewards are maximized by a neural network
- 7:45Deeper networks approximate discontinuous
- 10:31Neural network approximates discounted
- 13:16How does the neural network architecture look?
- 15:51How to update the neural network with Amazon SageMaker
- 18:37Recap: AWS DeepRacer learns to drive fast
- 20:09CAPCOM builds fun games faster with managed services on AWS
- 21:59MathWorks simulation environments enable industrial workflows
- 25:39Announcing new 2021 League features
related talks
24mGet Rolling on Machine Learning
29mGet rolling with ML on AWS DeepRacer
37mAWS DeepRacer: Train, Evaluate, and Tune
27mGet rolling with machine learning on AWS DeepRacer
37mAWS DeepRacer:train, Evaluate, and Tune Your Reinforcement Learning Model
42mFrom Zero to Machine Learning in 45 Minutes
28mShift your ML model into overdrive with AWS DeepRacer analysis tools
37mAccelerate your team’s machine learning journey with DeepRacer